diff mbox series

[1/4] X86: Update mmu_cr4_features during feature identification

Message ID 20200617190757.27081-2-john.s.andersen@intel.com (mailing list archive)
State New, archived
Headers show
Series Paravirtualized Control Register pinning | expand

Commit Message

Andersen, John June 17, 2020, 7:07 p.m. UTC
In identify_cpu when setting up SMEP/SMAP/UMIP call
cr4_set_bits_and_update_boot instead of cr4_set_bits. This ensures that
mmu_cr4_features contains those bits, and does not disable those
protections when in hibernation asm.

setup_arch updates mmu_cr4_features to save what identified features are
supported for later use in hibernation asm when cr4 needs to be modified
to toggle PGE. cr4 writes happen in restore_image and restore_registers.
setup_arch occurs before identify_cpu, this leads to mmu_cr4_features
not containing some of the cr4 features which were enabled via
identify_cpu when hibernation asm is executed.

On CPU bringup when cr4_set_bits_and_update_boot is called
mmu_cr4_features will now be written to. For the boot CPU,
the __ro_after_init on mmu_cr4_features does not cause a fault. However,
__ro_after_init was removed due to it triggering faults on non-boot
CPUs.

Signed-off-by: John Andersen <john.s.andersen@intel.com>
---
 arch/x86/kernel/cpu/common.c | 6 +++---
 arch/x86/kernel/setup.c      | 4 ++--
 2 files changed, 5 insertions(+), 5 deletions(-)

Comments

Dave Hansen June 18, 2020, 2:09 p.m. UTC | #1
Overall this looks pretty good.  For all the maintainers on cc, we try
to do internal reviews of these before they're submitted.  This one got
missed, sorry about that.

On 6/17/20 12:07 PM, John Andersen wrote:
> In identify_cpu when setting up SMEP/SMAP/UMIP call
> cr4_set_bits_and_update_boot instead of cr4_set_bits. This ensures that
> mmu_cr4_features contains those bits, and does not disable those
> protections when in hibernation asm.

When I'm writing comments, I try to use parenthesis for functions(),
which leaves variable_names plain.

I also try not to dive directly into the function names.  This
description assumes that the reader knows the subtle difference between
cr4_set_bits_and_update_boot() and of cr4_set_bits().  A sentence or two
of background here can save a reviewer a dive into the source code.

> setup_arch updates mmu_cr4_features to save what identified features are
> supported for later use in hibernation asm when cr4 needs to be modified
> to toggle PGE. cr4 writes happen in restore_image and restore_registers.
> setup_arch occurs before identify_cpu, this leads to mmu_cr4_features
> not containing some of the cr4 features which were enabled via
> identify_cpu when hibernation asm is executed.

This fails to address the bigger picture.  I assume you end up wanting
this because without it hibernation is not compatible with CR pinning.
Shouldn't that be mentioned?

I also wonder why we even need two classes of cr4_set_bits().  Are there
features we *want* to disable before entering the hibernation assembly?

For instance, why not leave MCE enabled in there?  What about PCIDs or
OSPKE?  Does it hurt?

> On CPU bringup when cr4_set_bits_and_update_boot is called 
> mmu_cr4_features will now be written to. For the boot CPU, the
> __ro_after_init on mmu_cr4_features does not cause a fault. However, 
> __ro_after_init was removed due to it triggering faults on non-boot 
> CPUs
Before this patch, cr4_set_bits_and_update_boot() was only ever called
during init.  But, after this patch, it gets called later in boot and
causes problems.  We're surely not making _real_ updates to it, right?
In that case the writes are superfluous and we would be better off just
not writing to it (and retaining __ro_after_init) rather than allowing
superfluous writes.
diff mbox series

Patch

diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c
index d07809286b95..921e67086a00 100644
--- a/arch/x86/kernel/cpu/common.c
+++ b/arch/x86/kernel/cpu/common.c
@@ -297,7 +297,7 @@  __setup("nosmep", setup_disable_smep);
 static __always_inline void setup_smep(struct cpuinfo_x86 *c)
 {
 	if (cpu_has(c, X86_FEATURE_SMEP))
-		cr4_set_bits(X86_CR4_SMEP);
+		cr4_set_bits_and_update_boot(X86_CR4_SMEP);
 }
 
 static __init int setup_disable_smap(char *arg)
@@ -316,7 +316,7 @@  static __always_inline void setup_smap(struct cpuinfo_x86 *c)
 
 	if (cpu_has(c, X86_FEATURE_SMAP)) {
 #ifdef CONFIG_X86_SMAP
-		cr4_set_bits(X86_CR4_SMAP);
+		cr4_set_bits_and_update_boot(X86_CR4_SMAP);
 #else
 		cr4_clear_bits(X86_CR4_SMAP);
 #endif
@@ -333,7 +333,7 @@  static __always_inline void setup_umip(struct cpuinfo_x86 *c)
 	if (!cpu_has(c, X86_FEATURE_UMIP))
 		goto out;
 
-	cr4_set_bits(X86_CR4_UMIP);
+	cr4_set_bits_and_update_boot(X86_CR4_UMIP);
 
 	pr_info_once("x86/cpu: User Mode Instruction Prevention (UMIP) activated\n");
 
diff --git a/arch/x86/kernel/setup.c b/arch/x86/kernel/setup.c
index a3767e74c758..d9c678b37a9b 100644
--- a/arch/x86/kernel/setup.c
+++ b/arch/x86/kernel/setup.c
@@ -138,9 +138,9 @@  EXPORT_SYMBOL(boot_cpu_data);
 
 
 #if !defined(CONFIG_X86_PAE) || defined(CONFIG_X86_64)
-__visible unsigned long mmu_cr4_features __ro_after_init;
+__visible unsigned long mmu_cr4_features;
 #else
-__visible unsigned long mmu_cr4_features __ro_after_init = X86_CR4_PAE;
+__visible unsigned long mmu_cr4_features = X86_CR4_PAE;
 #endif
 
 /* Boot loader ID and version as integers, for the benefit of proc_dointvec */