Commit dd78b97367 ("x86, boot: Move CPU
flags out of cpucheck") introduced ambiguous inline asm in the
has_eflag() function. In 16-bit mode want the instruction to be
'pushfl', but we just say 'pushf' and hope the compiler does what we
wanted.
When building with 'clang -m16', it won't, because clang doesn't use
the horrid '.code16gcc' hack that even 'gcc -m16' uses internally.
Say what we mean and don't make the compiler make assumptions.
[ hpa: ideally we would be able to use the gcc %zN construct here, but
that is broken for 64-bit integers in gcc < 4.5.
The code with plain "pushf/popf" is fine for 32- or 64-bit mode, but
not for 16-bit mode; in 16-bit mode those are 16-bit instructions in
.code16 mode, and 32-bit instructions in .code16gcc mode. ]
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Link: http://lkml.kernel.org/r/1391079628.26079.82.camel@shinybook.infradead.org
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
It looks like GCC will always emit an object that is marked with an
explicit section, although the documentation doesn't say that and we
possibly shouldn't be relying on it.
Clang does *not* do so, so add __attribute__((used)) to make sure.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Link: http://lkml.kernel.org/r/1389180083-23249-2-git-send-email-David.Woodhouse@intel.com
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Pull x86 kernel address space randomization support from Peter Anvin:
"This enables kernel address space randomization for x86"
* 'x86-kaslr-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86, kaslr: Clarify RANDOMIZE_BASE_MAX_OFFSET
x86, kaslr: Remove unused including <linux/version.h>
x86, kaslr: Use char array to gain sizeof sanity
x86, kaslr: Add a circular multiply for better bit diffusion
x86, kaslr: Mix entropy sources together as needed
x86/relocs: Add percpu fixup for GNU ld 2.23
x86, boot: Rename get_flags() and check_flags() to *_cpuflags()
x86, kaslr: Raise the maximum virtual address to -1 GiB on x86_64
x86, kaslr: Report kernel offset on panic
x86, kaslr: Select random position from e820 maps
x86, kaslr: Provide randomness functions
x86, kaslr: Return location from decompress_kernel
x86, boot: Move CPU flags out of cpucheck
x86, relocs: Add more per-cpu gold special cases
Pull x86 EFI changes from Ingo Molnar:
"This consists of two main parts:
- New static EFI runtime services virtual mapping layout which is
groundwork for kexec support on EFI (Borislav Petkov)
- EFI kexec support itself (Dave Young)"
* 'x86-efi-kexec-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (24 commits)
x86/efi: parse_efi_setup() build fix
x86: ksysfs.c build fix
x86/efi: Delete superfluous global variables
x86: Reserve setup_data ranges late after parsing memmap cmdline
x86: Export x86 boot_params to sysfs
x86: Add xloadflags bit for EFI runtime support on kexec
x86/efi: Pass necessary EFI data for kexec via setup_data
efi: Export EFI runtime memory mapping to sysfs
efi: Export more EFI table variables to sysfs
x86/efi: Cleanup efi_enter_virtual_mode() function
x86/efi: Fix off-by-one bug in EFI Boot Services reservation
x86/efi: Add a wrapper function efi_map_region_fixed()
x86/efi: Remove unused variables in __map_region()
x86/efi: Check krealloc return value
x86/efi: Runtime services virtual mapping
x86/mm/cpa: Map in an arbitrary pgd
x86/mm/pageattr: Add last levels of error path
x86/mm/pageattr: Add a PUD error unwinding path
x86/mm/pageattr: Add a PTE pagetable populating function
x86/mm/pageattr: Add a PMD pagetable populating function
...
The .inittext section tries to aggregate all functions which are
needed to get a message out in the case of a load failure. However,
putchar() uses intcall(), so intcall() should be in the .inittext
section.
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Link: http://lkml.kernel.org/n/tip-twxm8igouzbmsklmf6lfyq0w@git.kernel.org
This reverts commit 28b48688 ("x86, boot: use .code16gcc instead
of .code16").
Versions of binutils older than 2.16 are already not working, so this
workaround is no longer necessary either. At the same time, some of
the transformations that .code16gcc does can be *extremely*
counterintuitive to a human programmer.
[ hpa: folded ret -> retl and call -> calll fixes from followup patch ]
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Link: http://lkml.kernel.org/r/1388788242.2391.75.camel@shinybook.infradead.org
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Old kexec-tools can not load new kernels. The reason is kexec-tools does
not fill efi_info in x86 setup header previously, thus EFI failed to
initialize. In new kexec-tools it will by default to fill efi_info and
pass other EFI required infomation to 2nd kernel so kexec kernel EFI
initialization can succeed finally.
To prevent from breaking userspace, add a new xloadflags bit so
kexec-tools can check the flag and switch to old logic.
Signed-off-by: Dave Young <dyoung@redhat.com>
Acked-by: Borislav Petkov <bp@suse.de>
Tested-by: Toshi Kani <toshi.kani@hp.com>
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
In checkin
5551a34e5a x86-64, build: Always pass in -mno-sse
we unconditionally added -mno-sse to the main build, to keep newer
compilers from generating SSE instructions from autovectorization.
However, this did not extend to the special environments
(arch/x86/boot, arch/x86/boot/compressed, and arch/x86/realmode/rm).
Add -mno-sse to the compiler command line for these environments, and
add -mno-mmx to all the environments as well, as we don't want a
compiler to generate MMX code either.
This patch also removes a $(cc-option) call for -m32, since we have
long since stopped supporting compilers too old for the -m32 option,
and in fact hardcode it in other places in the Makefiles.
Reported-by: Kevin B. Smith <kevin.b.smith@intel.com>
Cc: Sunil K. Pandey <sunil.k.pandey@intel.com>
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Cc: H. J. Lu <hjl.tools@gmail.com>
Link: http://lkml.kernel.org/n/tip-j21wzqv790q834n7yc6g80j1@git.kernel.org
Cc: <stable@vger.kernel.org> # build fix only
The build_str needs to be char [] not char * for the sizeof() to report
the string length.
Reported-by: Mathias Krause <minipli@googlemail.com>
Signed-off-by: Kees Cook <keescook@chromium.org>
Link: http://lkml.kernel.org/r/20131112165607.GA5921@www.outflux.net
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
If we don't have RDRAND (in which case nothing else *should* matter),
most sources have a highly biased entropy distribution. Use a
circular multiply to diffuse the entropic bits. A circular multiply
is a good operation for this: it is cheap on standard hardware and
because it is symmetric (unlike an ordinary multiply) it doesn't
introduce its own bias.
Cc: Kees Cook <keescook@chromium.org>
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Link: http://lkml.kernel.org/r/20131111222839.GA28616@www.outflux.net
Depending on availability, mix the RDRAND and RDTSC entropy together with
XOR. Only when neither is available should the i8254 be used. Update
the Kconfig documentation to reflect this. Additionally, since bits
used for entropy is masked elsewhere, drop the needless masking in
the get_random_long(). Similarly, use the entire TSC, not just the low
32 bits.
Finally, to improve the starting entropy, do a simple hashing of a
build-time versions string and the boot-time boot_params structure for
some additional level of unpredictability.
Signed-off-by: Kees Cook <keescook@chromium.org>
Link: http://lkml.kernel.org/r/20131111222839.GA28616@www.outflux.net
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Pull x86 EFI changes from Ingo Molnar:
"Main changes:
- Add support for earlyprintk=efi which uses the EFI framebuffer.
Very useful for debugging boot problems.
- EFI stub support for large memory maps (more than 128 entries)
- EFI ARM support - this was mostly done by generalizing x86 <-> ARM
platform differences, such as by moving x86 EFI code into
drivers/firmware/efi/ and sharing it with ARM.
- Documentation updates
- misc fixes"
* 'x86-efi-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (26 commits)
x86/efi: Add EFI framebuffer earlyprintk support
boot, efi: Remove redundant memset()
x86/efi: Fix config_table_type array termination
x86 efi: bugfix interrupt disabling sequence
x86: EFI stub support for large memory maps
efi: resolve warnings found on ARM compile
efi: Fix types in EFI calls to match EFI function definitions.
efi: Renames in handle_cmdline_files() to complete generalization.
efi: Generalize handle_ramdisks() and rename to handle_cmdline_files().
efi: Allow efi_free() to be called with size of 0
efi: use efi_get_memory_map() to get final map for x86
efi: generalize efi_get_memory_map()
efi: Rename __get_map() to efi_get_memory_map()
efi: Move unicode to ASCII conversion to shared function.
efi: Generalize relocate_kernel() for use by other architectures.
efi: Move relocate_kernel() to shared file.
efi: Enforce minimum alignment of 1 page on allocations.
efi: Rename memory allocation/free functions
efi: Add system table pointer argument to shared functions.
efi: Move common EFI stub code from x86 arch code to common location
...
Pull x86 cleanups from Ingo Molnar:
"Two small cleanups"
* 'x86-cleanups-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86, msr: Use file_inode(), not f_mapping->host
x86: mkpiggy.c: Explicitly close the output file
When a function is used in more than one file it may not be possible
to immediately tell from context what the intended meaning is. As
such, it is more important that the naming be self-evident. Thus,
change get_flags() to get_cpuflags().
For consistency, change check_flags() to check_cpuflags() even though
it is only used in cpucheck.c.
Link: http://lkml.kernel.org/r/1381450698-28710-2-git-send-email-keescook@chromium.org
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Counts available alignment positions across all e820 maps, and chooses
one randomly for the new kernel base address, making sure not to collide
with unsafe memory areas.
Signed-off-by: Kees Cook <keescook@chromium.org>
Link: http://lkml.kernel.org/r/1381450698-28710-5-git-send-email-keescook@chromium.org
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Adds potential sources of randomness: RDRAND, RDTSC, or the i8254.
This moves the pre-alternatives inline rdrand function into the header so
both pieces of code can use it. Availability of RDRAND is then controlled
by CONFIG_ARCH_RANDOM, if someone wants to disable it even for kASLR.
Signed-off-by: Kees Cook <keescook@chromium.org>
Link: http://lkml.kernel.org/r/1381450698-28710-4-git-send-email-keescook@chromium.org
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
This allows decompress_kernel to return a new location for the kernel to
be relocated to. Additionally, enforces CONFIG_PHYSICAL_START as the
minimum relocation position when building with CONFIG_RELOCATABLE.
With CONFIG_RANDOMIZE_BASE set, the choose_kernel_location routine
will select a new location to decompress the kernel, though here it is
presently a no-op. The kernel command line option "nokaslr" is introduced
to bypass these routines.
Signed-off-by: Kees Cook <keescook@chromium.org>
Link: http://lkml.kernel.org/r/1381450698-28710-3-git-send-email-keescook@chromium.org
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Refactor the CPU flags handling out of the cpucheck routines so that
they can be reused by the future ASLR routines (in order to detect CPU
features like RDRAND and RDTSC).
This reworks has_eflag() and has_fpu() to be used on both 32-bit and
64-bit, and refactors the calls to cpuid to make them PIC-safe on 32-bit.
Signed-off-by: Kees Cook <keescook@chromium.org>
Link: http://lkml.kernel.org/r/1381450698-28710-2-git-send-email-keescook@chromium.org
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Even though the resource is released when the application is closed or
when returned from main function, modify the code to make it obvious,
and to keep static analysis tools from complaining.
Signed-off-by: Geyslan G. Bem <geyslan@gmail.com>
Link: http://lkml.kernel.org/r/1381184219-10985-1-git-send-email-geyslan@gmail.com
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
The problem in efi_main was that the idt was cleared before the
interrupts were disabled.
The UEFI spec states that interrupts aren't used so this shouldn't be
too much of a problem. Peripherals however don't necessarily know about
this and thus might cause interrupts to happen anyway. Even if
ExitBootServices() has been called.
This means there is a risk of an interrupt being triggered while the IDT
register is nullified and the interrupt bit hasn't been cleared,
allowing for a triple fault.
This patch disables the interrupt flag, while leaving the existing IDT
in place. The CPU won't care about the IDT at all as long as the
interrupt bit is off, so it's safe to leave it in place as nothing will
ever happen to it.
[ Removed the now unused 'idt' variable - Matt ]
Signed-off-by: Bart Kuivenhoven <bemk@redhat.com>
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
This patch fixes a problem with EFI memory maps larger than 128 entries
when booting using the EFI stub, which results in overflowing e820_map
in boot_params and an eventual halt when checking the map size in
sanitize_e820_map().
If the number of map entries is greater than what can fit in e820_map,
add the extra entries to the setup_data list using type SETUP_E820_EXT.
These extra entries are then picked up when the setup_data list is
parsed in parse_e820_ext().
Signed-off-by: Linn Crosetto <linn@hp.com>
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
When building on x86, the final image building step always emits stats
to stderr, even though this information is neither a warning nor an error:
BUILD arch/x86/boot/bzImage
Setup is 16188 bytes (padded to 16384 bytes).
System is 6368 kB
CRC cbe50c61
Validating automated builds would be cleaner if stderr did not have to
filter out these lines. Instead, change how tools/build is called, and
make the zoffset header unconditional, and write to a specified file
instead of to stdout, which can then be used for statistics, leaving
stderr open for legitimate warnings and errors, like the output from
die().
Signed-off-by: Kees Cook <keescook@chromium.org>
Link: http://lkml.kernel.org/r/20130906181532.GA31260@www.outflux.net
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
The handle_cmdline_files now takes the option to handle as a string,
and returns the loaded data through parameters, rather than taking
an x86 specific setup_header structure. For ARM, this will be used
to load a device tree blob in addition to initrd images.
Signed-off-by: Roy Franz <roy.franz@linaro.org>
Acked-by: Mark Salter <msalter@redhat.com>
Reviewed-by: Grant Likely <grant.likely@linaro.org>
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Make efi_free() safely callable with size of 0, similar to free() being
callable with NULL pointers, and do nothing in that case.
Remove size checks that this makes redundant. This also avoids some
size checks in the ARM EFI stub code that will be added as well.
Signed-off-by: Roy Franz <roy.franz@linaro.org>
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Replace the open-coded memory map getting with the
efi_get_memory_map() that is now general enough to use.
Signed-off-by: Roy Franz <roy.franz@linaro.org>
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Move the open-coded conversion to a shared function for
use by all architectures. Change the allocation to prefer
a high address for ARM, as this is required to avoid conflicts
with reserved regions in low memory. We don't know the specifics
of these regions until after we process the command line and
device tree.
Signed-off-by: Roy Franz <roy.franz@linaro.org>
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Rename relocate_kernel() to efi_relocate_kernel(), and take
parameters rather than x86 specific structure. Add max_addr
argument as for ARM we have some address constraints that we
need to enforce when relocating the kernel. Add alloc_size
parameter for use by ARM64 which uses an uncompressed kernel,
and needs to allocate space for BSS.
Signed-off-by: Roy Franz <roy.franz@linaro.org>
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
The relocate_kernel() function will be generalized and used
by all architectures, as they all have similar requirements.
Signed-off-by: Roy Franz <roy.franz@linaro.org>
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Rename them to be more similar, as low_free() could be used to free
memory allocated by both high_alloc() and low_alloc().
high_alloc() -> efi_high_alloc()
low_alloc() -> efi_low_alloc()
low_free() -> efi_free()
Signed-off-by: Roy Franz <roy.franz@linaro.org>
Acked-by: Mark Salter <msalter@redhat.com>
Reviewed-by: Grant Likely <grant.likely@linaro.org>
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Add system table pointer argument to shared EFI stub related functions
so they no longer use a global system table pointer as they did when part
of eboot.c. For the ARM EFI stub this allows us to avoid global
variables completely and thereby not have to deal with GOT fixups.
Not having the EFI stub fixup its GOT, which is shared with the
decompressor, simplifies the relocating of the zImage to a
bootable address.
Signed-off-by: Roy Franz <roy.franz@linaro.org>
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
No code changes made, just moving functions and #define from x86 arch
directory to common location. Code is shared using #include, similar
to how decompression code is shared among architectures.
Signed-off-by: Roy Franz <roy.franz@linaro.org>
Acked-by: Mark Salter <msalter@redhat.com>
Reviewed-by: Grant Likely <grant.likely@linaro.org>
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
The x86/AMD64 EFI stubs must use a call wrapper to convert between
the Linux and EFI ABIs, so void pointers are sufficient. For ARM,
the ABIs are compatible, so we can directly invoke the function
pointers. The functions that are used by the ARM stub are updated
to match the EFI definitions.
Also add some EFI types used by EFI functions.
Signed-off-by: Roy Franz <roy.franz@linaro.org>
Acked-by: Mark Salter <msalter@redhat.com>
Reviewed-by: Grant Likely <grant.likely@linaro.org>
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Pull x86 relocation changes from Ingo Molnar:
"This tree contains a single change, ELF relocation handling in C - one
of the kernel randomization patches that makes sense even without
randomization present upstream"
* 'x86-kaslr-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86, relocs: Move ELF relocation handling to C
Pull tiny x86 boot cleanups from Ingo Molnar.
* 'x86-boot-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86/boot: Fix a sanity check in printf.c
* 'x86-cleanups-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86, boot: Fix warning due to undeclared strlen()
Prior to 9b706aee7d ("x86: trivial printk optimizations") this was
36 because it had 26 characters and 10 digits but now it's just
16 hex digits so the sanity check needs updated.
This function is always called with a valid "base" so it doesn't
make a difference to how the kernel works, it's just a cleanup.
Reported-by: Alexey Petrenko <alexey.petrenko@oracle.com>
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Below is a patch that fixes sparse error
"arch/x86/boot/string.c:119:8: warning: symbol 'strlen' was not
declared." by declaring it in arch/x86/boot/boot.h.
Signed-off-by: Fred Chen <fchen@linux.vnet.ibm.com>
Link: http://lkml.kernel.org/r/1376417580-11554-1-git-send-email-fchen@linux.vnet.ibm.com
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Moves the relocation handling into C, after decompression. This requires
that the decompressed size is passed to the decompression routine as
well so that relocations can be found. Only kernels that need relocation
support will use the code (currently just x86_32), but this is laying
the ground work for 64-bit using it in support of KASLR.
Based on work by Neill Clift and Michael Davidson.
Signed-off-by: Kees Cook <keescook@chromium.org>
Link: http://lkml.kernel.org/r/20130708161517.GA4832@www.outflux.net
Acked-by: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Pull x86 mm changes from Ingo Molnar:
"Misc improvements:
- Fix /proc/mtrr reporting
- Fix ioremap printout
- Remove the unused pvclock fixmap entry on 32-bit
- misc cleanups"
* 'x86-mm-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86/ioremap: Correct function name output
x86: Fix /proc/mtrr with base/size more than 44bits
ix86: Don't waste fixmap entries
x86/mm: Drop unneeded include <asm/*pgtable, page*_types.h>
x86_64: Correct phys_addr in cleanup_highmap comment
Pull x86 EFI changes from Ingo Molnar:
"Two fixes that should in principle increase robustness of our
interaction with the EFI firmware, and a cleanup"
* 'x86-efi-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86, efi: retry ExitBootServices() on failure
efi: Convert runtime services function ptrs
UEFI: Don't pass boot services regions to SetVirtualAddressMap()
ExitBootServices is absolutely supposed to return a failure if any
ExitBootServices event handler changes the memory map. Basically the
get_map loop should run again if ExitBootServices returns an error the
first time. I would say it would be fair that if ExitBootServices gives
an error the second time then Linux would be fine in returning control
back to BIOS.
The second change is the following line:
again:
size += sizeof(*mem_map) * 2;
Originally you were incrementing it by the size of one memory map entry.
The issue here is all related to the low_alloc routine you are using.
In this routine you are making allocations to get the memory map itself.
Doing this allocation or allocations can affect the memory map by more
than one record.
[ mfleming - changelog, code style ]
Signed-off-by: Zach Bobroff <zacharyb@ami.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
This patch reworks the UEFI anti-bricking code, including an effective
reversion of cc5a080c and 31ff2f20. It turns out that calling
QueryVariableInfo() from boot services results in some firmware
implementations jumping to physical addresses even after entering virtual
mode, so until we have 1:1 mappings for UEFI runtime space this isn't
going to work so well.
Reverting these gets us back to the situation where we'd refuse to create
variables on some systems because they classify deleted variables as "used"
until the firmware triggers a garbage collection run, which they won't do
until they reach a lower threshold. This results in it being impossible to
install a bootloader, which is unhelpful.
Feedback from Samsung indicates that the firmware doesn't need more than
5KB of storage space for its own purposes, so that seems like a reasonable
threshold. However, there's still no guarantee that a platform will attempt
garbage collection merely because it drops below this threshold. It seems
that this is often only triggered if an attempt to write generates a
genuine EFI_OUT_OF_RESOURCES error. We can force that by attempting to
create a variable larger than the remaining space. This should fail, but if
it somehow succeeds we can then immediately delete it.
I've tested this on the UEFI machines I have available, but I don't have
a Samsung and so can't verify that it avoids the bricking problem.
Signed-off-by: Matthew Garrett <matthew.garrett@nebula.com>
Signed-off-by: Lee, Chun-Y <jlee@suse.com> [ dummy variable cleanup ]
Cc: <stable@vger.kernel.org>
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
arch/x86/boot/compressed/head_64.S includes <asm/pgtable_types.h> and
<asm/page_types.h> but it doesn't look like it needs them. So remove them.
Signed-off-by: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Link: http://lkml.kernel.org/r/5191FAE2.4020403@cn.fujitsu.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
