x86, 32-bit: trim memory not covered by wb mtrrs

On some machines, buggy BIOSes don't properly setup WB MTRRs to cover all
available RAM, meaning the last few megs (or even gigs) of memory will be
marked uncached.  Since Linux tends to allocate from high memory addresses
first, this causes the machine to be unusably slow as soon as the kernel
starts really using memory (i.e.  right around init time).

This patch works around the problem by scanning the MTRRs at boot and
figuring out whether the current end_pfn value (setup by early e820 code)
goes beyond the highest WB MTRR range, and if so, trimming it to match.  A
fairly obnoxious KERN_WARNING is printed too, letting the user know that
not all of their memory is available due to a likely BIOS bug.

Something similar could be done on i386 if needed, but the boot ordering
would be slightly different, since the MTRR code on i386 depends on the
boot_cpu_data structure being setup.

This patch fixes a bug in the last patch that caused the code to run on
non-Intel machines (AMD machines apparently don't need it and it's untested
on other non-Intel machines, so best keep it off).

Further enhancements and fixes from:

  Yinghai Lu <Yinghai.Lu@Sun.COM>
  Andi Kleen <ak@suse.de>

Signed-off-by: Jesse Barnes <jesse.barnes@intel.com>
Tested-by: Justin Piszcz <jpiszcz@lucidpixels.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Yinghai Lu <yhlu.kernel@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>

Documentation/kernel-parameters.txt

@@ -570,6 +570,12 @@ and is between 256 and 4096 characters. It is defined in the file
 			See drivers/char/README.epca and
 			Documentation/digiepca.txt.
 
+	disable_mtrr_trim [X86-64, Intel only]
+			By default the kernel will trim any uncacheable
+			memory out of your available memory pool based on
+			MTRR settings.  This parameter disables that behavior,
+			possibly causing your machine to run very slowly.
+
 	dmasound=	[HW,OSS] Sound subsystem buffers
 
 	dscc4.setup=	[NET]

arch/x86/kernel/bugs_64.c

@@ -13,7 +13,6 @@
 void __init check_bugs(void)
 {
 	identify_cpu(&boot_cpu_data);
-	mtrr_bp_init();
 #if !defined(CONFIG_SMP)
 	printk("CPU: ");
 	print_cpu_info(&boot_cpu_data);

arch/x86/kernel/cpu/mtrr/generic.c

@@ -14,7 +14,7 @@
 #include "mtrr.h"
 
 struct mtrr_state {
-	struct mtrr_var_range *var_ranges;
+	struct mtrr_var_range var_ranges[MAX_VAR_RANGES];
 	mtrr_type fixed_ranges[NUM_FIXED_RANGES];
 	unsigned char enabled;
 	unsigned char have_fixed;
@@ -86,12 +86,6 @@ void __init get_mtrr_state(void)
 	struct mtrr_var_range *vrs;
 	unsigned lo, dummy;
 
-	if (!mtrr_state.var_ranges) {
-		mtrr_state.var_ranges = kmalloc(num_var_ranges * sizeof (struct mtrr_var_range), 
-						GFP_KERNEL);
-		if (!mtrr_state.var_ranges)
-			return;
-	} 
 	vrs = mtrr_state.var_ranges;
 
 	rdmsr(MTRRcap_MSR, lo, dummy);

arch/x86/kernel/cpu/mtrr/if.c

@@ -11,10 +11,6 @@
 #include <asm/mtrr.h>
 #include "mtrr.h"
 
-/* RED-PEN: this is accessed without any locking */
-extern unsigned int *usage_table;
-
-
 #define FILE_FCOUNT(f) (((struct seq_file *)((f)->private_data))->private)
 
 static const char *const mtrr_strings[MTRR_NUM_TYPES] =
@@ -397,7 +393,7 @@ static int mtrr_seq_show(struct seq_file *seq, void *offset)
 	for (i = 0; i < max; i++) {
 		mtrr_if->get(i, &base, &size, &type);
 		if (size == 0)
-			usage_table[i] = 0;
+			mtrr_usage_table[i] = 0;
 		else {
 			if (size < (0x100000 >> PAGE_SHIFT)) {
 				/* less than 1MB */
@@ -411,7 +407,7 @@ static int mtrr_seq_show(struct seq_file *seq, void *offset)
 			len += seq_printf(seq, 
 				   "reg%02i: base=0x%05lx000 (%4luMB), size=%4lu%cB: %s, count=%d\n",
 			     i, base, base >> (20 - PAGE_SHIFT), size, factor,
-			     mtrr_attrib_to_str(type), usage_table[i]);
+			     mtrr_attrib_to_str(type), mtrr_usage_table[i]);
 		}
 	}
 	return 0;

arch/x86/kernel/cpu/mtrr/main.c

@@ -38,8 +38,8 @@
 #include <linux/cpu.h>
 #include <linux/mutex.h>
 
+#include <asm/e820.h>
 #include <asm/mtrr.h>
-
 #include <asm/uaccess.h>
 #include <asm/processor.h>
 #include <asm/msr.h>
@@ -47,7 +47,7 @@
 
 u32 num_var_ranges = 0;
 
-unsigned int *usage_table;
+unsigned int mtrr_usage_table[MAX_VAR_RANGES];
 static DEFINE_MUTEX(mtrr_mutex);
 
 u64 size_or_mask, size_and_mask;
@@ -121,13 +121,8 @@ static void __init init_table(void)
 	int i, max;
 
 	max = num_var_ranges;
-	if ((usage_table = kmalloc(max * sizeof *usage_table, GFP_KERNEL))
-	    == NULL) {
-		printk(KERN_ERR "mtrr: could not allocate\n");
-		return;
-	}
 	for (i = 0; i < max; i++)
-		usage_table[i] = 1;
+		mtrr_usage_table[i] = 1;
 }
 
 struct set_mtrr_data {
@@ -383,7 +378,7 @@ int mtrr_add_page(unsigned long base, unsigned long size,
 			goto out;
 		}
 		if (increment)
-			++usage_table[i];
+			++mtrr_usage_table[i];
 		error = i;
 		goto out;
 	}
@@ -391,15 +386,15 @@ int mtrr_add_page(unsigned long base, unsigned long size,
 	i = mtrr_if->get_free_region(base, size, replace);
 	if (i >= 0) {
 		set_mtrr(i, base, size, type);
-		if (likely(replace < 0))
-			usage_table[i] = 1;
-		else {
-			usage_table[i] = usage_table[replace];
+		if (likely(replace < 0)) {
+			mtrr_usage_table[i] = 1;
+		} else {
+			mtrr_usage_table[i] = mtrr_usage_table[replace];
 			if (increment)
-				usage_table[i]++;
+				mtrr_usage_table[i]++;
 			if (unlikely(replace != i)) {
 				set_mtrr(replace, 0, 0, 0);
-				usage_table[replace] = 0;
+				mtrr_usage_table[replace] = 0;
 			}
 		}
 	} else
@@ -529,11 +524,11 @@ int mtrr_del_page(int reg, unsigned long base, unsigned long size)
 		printk(KERN_WARNING "mtrr: MTRR %d not used\n", reg);
 		goto out;
 	}
-	if (usage_table[reg] < 1) {
+	if (mtrr_usage_table[reg] < 1) {
 		printk(KERN_WARNING "mtrr: reg: %d has count=0\n", reg);
 		goto out;
 	}
-	if (--usage_table[reg] < 1)
+	if (--mtrr_usage_table[reg] < 1)
 		set_mtrr(reg, 0, 0, 0);
 	error = reg;
  out:
@@ -593,16 +588,11 @@ struct mtrr_value {
 	unsigned long	lsize;
 };
 
-static struct mtrr_value * mtrr_state;
+static struct mtrr_value mtrr_state[MAX_VAR_RANGES];
 
 static int mtrr_save(struct sys_device * sysdev, pm_message_t state)
 {
 	int i;
-	int size = num_var_ranges * sizeof(struct mtrr_value);
-
-	mtrr_state = kzalloc(size,GFP_ATOMIC);
-	if (!mtrr_state)
-		return -ENOMEM;
 
 	for (i = 0; i < num_var_ranges; i++) {
 		mtrr_if->get(i,
@@ -624,7 +614,6 @@ static int mtrr_restore(struct sys_device * sysdev)
 				 mtrr_state[i].lsize,
 				 mtrr_state[i].ltype);
 	}
-	kfree(mtrr_state);
 	return 0;
 }
 
@@ -635,6 +624,109 @@ static struct sysdev_driver mtrr_sysdev_driver = {
 	.resume		= mtrr_restore,
 };
 
+#ifdef CONFIG_X86_64
+static int disable_mtrr_trim;
+
+static int __init disable_mtrr_trim_setup(char *str)
+{
+	disable_mtrr_trim = 1;
+	return 0;
+}
+early_param("disable_mtrr_trim", disable_mtrr_trim_setup);
+
+/*
+ * Newer AMD K8s and later CPUs have a special magic MSR way to force WB
+ * for memory >4GB. Check for that here.
+ * Note this won't check if the MTRRs < 4GB where the magic bit doesn't
+ * apply to are wrong, but so far we don't know of any such case in the wild.
+ */
+#define Tom2Enabled (1U << 21)
+#define Tom2ForceMemTypeWB (1U << 22)
+
+static __init int amd_special_default_mtrr(unsigned long end_pfn)
+{
+	u32 l, h;
+
+	/* Doesn't apply to memory < 4GB */
+	if (end_pfn <= (0xffffffff >> PAGE_SHIFT))
+		return 0;
+	if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
+		return 0;
+	if (boot_cpu_data.x86 < 0xf || boot_cpu_data.x86 > 0x11)
+		return 0;
+	/* In case some hypervisor doesn't pass SYSCFG through */
+	if (rdmsr_safe(MSR_K8_SYSCFG, &l, &h) < 0)
+		return 0;
+	/*
+	 * Memory between 4GB and top of mem is forced WB by this magic bit.
+	 * Reserved before K8RevF, but should be zero there.
+	 */
+	if ((l & (Tom2Enabled | Tom2ForceMemTypeWB)) ==
+		 (Tom2Enabled | Tom2ForceMemTypeWB))
+		return 1;
+	return 0;
+}
+
+/**
+ * mtrr_trim_uncached_memory - trim RAM not covered by MTRRs
+ *
+ * Some buggy BIOSes don't setup the MTRRs properly for systems with certain
+ * memory configurations.  This routine checks that the highest MTRR matches
+ * the end of memory, to make sure the MTRRs having a write back type cover
+ * all of the memory the kernel is intending to use. If not, it'll trim any
+ * memory off the end by adjusting end_pfn, removing it from the kernel's
+ * allocation pools, warning the user with an obnoxious message.
+ */
+int __init mtrr_trim_uncached_memory(unsigned long end_pfn)
+{
+	unsigned long i, base, size, highest_addr = 0, def, dummy;
+	mtrr_type type;
+	u64 trim_start, trim_size;
+
+	/*
+	 * Make sure we only trim uncachable memory on machines that
+	 * support the Intel MTRR architecture:
+	 */
+	rdmsr(MTRRdefType_MSR, def, dummy);
+	def &= 0xff;
+	if (!is_cpu(INTEL) || disable_mtrr_trim || def != MTRR_TYPE_UNCACHABLE)
+		return 0;
+
+	/* Find highest cached pfn */
+	for (i = 0; i < num_var_ranges; i++) {
+		mtrr_if->get(i, &base, &size, &type);
+		if (type != MTRR_TYPE_WRBACK)
+			continue;
+		base <<= PAGE_SHIFT;
+		size <<= PAGE_SHIFT;
+		if (highest_addr < base + size)
+			highest_addr = base + size;
+	}
+
+	if (amd_special_default_mtrr(end_pfn))
+		return 0;
+
+	if ((highest_addr >> PAGE_SHIFT) < end_pfn) {
+		printk(KERN_WARNING "***************\n");
+		printk(KERN_WARNING "**** WARNING: likely BIOS bug\n");
+		printk(KERN_WARNING "**** MTRRs don't cover all of "
+		       "memory, trimmed %ld pages\n", end_pfn -
+		       (highest_addr >> PAGE_SHIFT));
+		printk(KERN_WARNING "***************\n");
+
+		printk(KERN_INFO "update e820 for mtrr\n");
+		trim_start = highest_addr;
+		trim_size = end_pfn;
+		trim_size <<= PAGE_SHIFT;
+		trim_size -= trim_start;
+		add_memory_region(trim_start, trim_size, E820_RESERVED);
+		update_e820();
+		return 1;
+	}
+
+	return 0;
+}
+#endif
 
 /**
  * mtrr_bp_init - initialize mtrrs on the boot CPU

arch/x86/kernel/cpu/mtrr/mtrr.h

@@ -12,6 +12,7 @@
 #define MTRRphysMask_MSR(reg) (0x200 + 2 * (reg) + 1)
 
 #define NUM_FIXED_RANGES 88
+#define MAX_VAR_RANGES 256
 #define MTRRfix64K_00000_MSR 0x250
 #define MTRRfix16K_80000_MSR 0x258
 #define MTRRfix16K_A0000_MSR 0x259
@@ -32,6 +33,8 @@
    an 8 bit field: */
 typedef u8 mtrr_type;
 
+extern unsigned int mtrr_usage_table[MAX_VAR_RANGES];
+
 struct mtrr_ops {
 	u32	vendor;
 	u32	use_intel_if;

arch/x86/kernel/setup_64.c

@@ -310,6 +310,13 @@ void __init setup_arch(char **cmdline_p)
 	 * we are rounding upwards:
 	 */
 	end_pfn = e820_end_of_ram();
+	/* update e820 for memory not covered by WB MTRRs */
+	mtrr_bp_init();
+	if (mtrr_trim_uncached_memory(end_pfn)) {
+		e820_register_active_regions(0, 0, -1UL);
+		end_pfn = e820_end_of_ram();
+	}
+
 	num_physpages = end_pfn;
 
 	check_efer();

include/asm-x86/mtrr.h

@@ -97,6 +97,7 @@ extern int mtrr_del_page (int reg, unsigned long base, unsigned long size);
 extern void mtrr_centaur_report_mcr(int mcr, u32 lo, u32 hi);
 extern void mtrr_ap_init(void);
 extern void mtrr_bp_init(void);
+extern int mtrr_trim_uncached_memory(unsigned long end_pfn);
 #  else
 #define mtrr_save_fixed_ranges(arg) do {} while (0)
 #define mtrr_save_state() do {} while (0)
@@ -120,7 +121,10 @@ static __inline__ int mtrr_del_page (int reg, unsigned long base,
 {
     return -ENODEV;
 }
-
+static inline int mtrr_trim_uncached_memory(unsigned long end_pfn)
+{
+	return 0;
+}
 static __inline__ void mtrr_centaur_report_mcr(int mcr, u32 lo, u32 hi) {;}
 
 #define mtrr_ap_init() do {} while (0)