If some config options are disabled during compile time, they still are enumerated in macros that use the x86_capability bitmask - cpu_has() or this_cpu_has().
The features are also visible in /proc/cpuinfo even though they are not enabled - which is contrary to what the documentation states about the file. Examples of such feature flags are lam, fred, sgx, ibrs_enhanced, split_lock_detect, user_shstk, avx_vnni and enqcmd.
Add a DISABLED_MASK() macro that returns 32 bit chunks of the disabled feature bits bitmask.
Initialize the cpu_caps_cleared and cpu_caps_set arrays with the contents of the disabled and required bitmasks respectively. Then let apply_forced_caps() clear/set these feature bits in the x86_capability.
Fixes: 6449dcb0cac7 ("x86: CPUID and CR3/CR4 flags for Linear Address Masking") Fixes: 51c158f7aacc ("x86/cpufeatures: Add the CPU feature bit for FRED") Fixes: 706d51681d63 ("x86/speculation: Support Enhanced IBRS on future CPUs") Fixes: e7b6385b01d8 ("x86/cpufeatures: Add Intel SGX hardware bits") Fixes: 6650cdd9a8cc ("x86/split_lock: Enable split lock detection by kernel") Fixes: 701fb66d576e ("x86/cpufeatures: Add CPU feature flags for shadow stacks") Fixes: ff4f82816dff ("x86/cpufeatures: Enumerate ENQCMD and ENQCMDS instructions") Signed-off-by: Maciej Wieczor-Retman maciej.wieczor-retman@intel.com Cc: stable@vger.kernel.org --- Changelog v2: - Redo the patch to utilize a more generic solution, not just fix the LAM and FRED feature bits. - Note more feature flags that shouldn't be present. - Add fixes and cc tags.
arch/x86/kernel/cpu/common.c | 12 ++++++++++++ arch/x86/tools/cpufeaturemasks.awk | 8 ++++++++ 2 files changed, 20 insertions(+)
diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c index 77afca95cced..ba8b5fba8552 100644 --- a/arch/x86/kernel/cpu/common.c +++ b/arch/x86/kernel/cpu/common.c @@ -1709,6 +1709,16 @@ static void __init cpu_parse_early_param(void) } }
+static __init void init_cpu_cap(struct cpuinfo_x86 *c) +{ + int i; + + for (i = 0; i < NCAPINTS; i++) { + cpu_caps_set[i] = REQUIRED_MASK(i); + cpu_caps_cleared[i] = DISABLED_MASK(i); + } +} + /* * Do minimum CPU detection early. * Fields really needed: vendor, cpuid_level, family, model, mask, @@ -1782,6 +1792,8 @@ static void __init early_identify_cpu(struct cpuinfo_x86 *c) if (!pgtable_l5_enabled()) setup_clear_cpu_cap(X86_FEATURE_LA57);
+ init_cpu_cap(c); + detect_nopl(); }
diff --git a/arch/x86/tools/cpufeaturemasks.awk b/arch/x86/tools/cpufeaturemasks.awk index 173d5bf2d999..2e2412f7681f 100755 --- a/arch/x86/tools/cpufeaturemasks.awk +++ b/arch/x86/tools/cpufeaturemasks.awk @@ -82,6 +82,14 @@ END { } printf " 0\t\\n"; printf "\t) & (1U << ((x) & 31)))\n\n"; + + printf "\n#define %s_MASK(x)\t\t\t\t\\n", s; + printf "\t((\t\t\t\t"; + for (i = 0; i < ncapints; i++) { + if (masks[i]) + printf "\t\t\\n\t\t(x) == %2d ? %s_MASK%d :", i, s, i; + } + printf " 0))\t\\n\n"; }
printf "#endif /* _ASM_X86_CPUFEATUREMASKS_H */\n";
On Wed, Jul 23, 2025 at 11:22:49AM +0200, Maciej Wieczor-Retman wrote:
If some config options are disabled during compile time, they still are enumerated in macros that use the x86_capability bitmask - cpu_has() or this_cpu_has().
The features are also visible in /proc/cpuinfo even though they are not enabled - which is contrary to what the documentation states about the file. Examples of such feature flags are lam, fred, sgx, ibrs_enhanced, split_lock_detect, user_shstk, avx_vnni and enqcmd.
Add a DISABLED_MASK() macro that returns 32 bit chunks of the disabled feature bits bitmask.
Initialize the cpu_caps_cleared and cpu_caps_set arrays with the contents of the disabled and required bitmasks respectively. Then let apply_forced_caps() clear/set these feature bits in the x86_capability.
Fixes: 6449dcb0cac7 ("x86: CPUID and CR3/CR4 flags for Linear Address Masking") Fixes: 51c158f7aacc ("x86/cpufeatures: Add the CPU feature bit for FRED") Fixes: 706d51681d63 ("x86/speculation: Support Enhanced IBRS on future CPUs") Fixes: e7b6385b01d8 ("x86/cpufeatures: Add Intel SGX hardware bits") Fixes: 6650cdd9a8cc ("x86/split_lock: Enable split lock detection by kernel") Fixes: 701fb66d576e ("x86/cpufeatures: Add CPU feature flags for shadow stacks") Fixes: ff4f82816dff ("x86/cpufeatures: Enumerate ENQCMD and ENQCMDS instructions")
That is fricken insane.
You are saying to people who backport stuff: This fixes a commit found in the following kernel releases: 6.4 6.9 3.16.68 4.4.180 4.9.137 4.14.81 4.18.19 4.19 5.11 5.7 6.6 5.10
You didn't even sort this in any sane order, how was it generated?
What in the world is anyone supposed to do with this?
If you were sent a patch with this in it, what would you think? What could you do with it?
Please be reasonable and consider us overworked stable maintainers and give us a chance to get things right. As it is, this just makes things worse...
greg k-h
On 2025-07-23 at 11:45:22 +0200, Greg KH wrote:
On Wed, Jul 23, 2025 at 11:22:49AM +0200, Maciej Wieczor-Retman wrote:
If some config options are disabled during compile time, they still are enumerated in macros that use the x86_capability bitmask - cpu_has() or this_cpu_has().
The features are also visible in /proc/cpuinfo even though they are not enabled - which is contrary to what the documentation states about the file. Examples of such feature flags are lam, fred, sgx, ibrs_enhanced, split_lock_detect, user_shstk, avx_vnni and enqcmd.
Add a DISABLED_MASK() macro that returns 32 bit chunks of the disabled feature bits bitmask.
Initialize the cpu_caps_cleared and cpu_caps_set arrays with the contents of the disabled and required bitmasks respectively. Then let apply_forced_caps() clear/set these feature bits in the x86_capability.
Fixes: 6449dcb0cac7 ("x86: CPUID and CR3/CR4 flags for Linear Address Masking") Fixes: 51c158f7aacc ("x86/cpufeatures: Add the CPU feature bit for FRED") Fixes: 706d51681d63 ("x86/speculation: Support Enhanced IBRS on future CPUs") Fixes: e7b6385b01d8 ("x86/cpufeatures: Add Intel SGX hardware bits") Fixes: 6650cdd9a8cc ("x86/split_lock: Enable split lock detection by kernel") Fixes: 701fb66d576e ("x86/cpufeatures: Add CPU feature flags for shadow stacks") Fixes: ff4f82816dff ("x86/cpufeatures: Enumerate ENQCMD and ENQCMDS instructions")
That is fricken insane.
You are saying to people who backport stuff: This fixes a commit found in the following kernel releases: 6.4 6.9 3.16.68 4.4.180 4.9.137 4.14.81 4.18.19 4.19 5.11 5.7 6.6 5.10
You didn't even sort this in any sane order, how was it generated?
What in the world is anyone supposed to do with this?
If you were sent a patch with this in it, what would you think? What could you do with it?
Please be reasonable and consider us overworked stable maintainers and give us a chance to get things right. As it is, this just makes things worse...
greg k-h
Sorry, I certainly didn't want to add you more work.
I noted down which features are present in the x86_capability bitmask while they're not compiled into the kernel. Then I noted down which commits added these feature flags. So I suppose the order is from least to most significant feature bit, which now I realize doesn't help much in backporting, again sorry.
Would a more fitting Fixes: commit be the one that changed how the feature flags are used? At some point docs started stating to have them set only when features are COMPILED & HARDWARE-SUPPORTED.
On Wed, Jul 23, 2025 at 01:46:44PM +0200, Maciej Wieczor-Retman wrote:
On 2025-07-23 at 11:45:22 +0200, Greg KH wrote:
On Wed, Jul 23, 2025 at 11:22:49AM +0200, Maciej Wieczor-Retman wrote:
If some config options are disabled during compile time, they still are enumerated in macros that use the x86_capability bitmask - cpu_has() or this_cpu_has().
The features are also visible in /proc/cpuinfo even though they are not enabled - which is contrary to what the documentation states about the file. Examples of such feature flags are lam, fred, sgx, ibrs_enhanced, split_lock_detect, user_shstk, avx_vnni and enqcmd.
Add a DISABLED_MASK() macro that returns 32 bit chunks of the disabled feature bits bitmask.
Initialize the cpu_caps_cleared and cpu_caps_set arrays with the contents of the disabled and required bitmasks respectively. Then let apply_forced_caps() clear/set these feature bits in the x86_capability.
Fixes: 6449dcb0cac7 ("x86: CPUID and CR3/CR4 flags for Linear Address Masking") Fixes: 51c158f7aacc ("x86/cpufeatures: Add the CPU feature bit for FRED") Fixes: 706d51681d63 ("x86/speculation: Support Enhanced IBRS on future CPUs") Fixes: e7b6385b01d8 ("x86/cpufeatures: Add Intel SGX hardware bits") Fixes: 6650cdd9a8cc ("x86/split_lock: Enable split lock detection by kernel") Fixes: 701fb66d576e ("x86/cpufeatures: Add CPU feature flags for shadow stacks") Fixes: ff4f82816dff ("x86/cpufeatures: Enumerate ENQCMD and ENQCMDS instructions")
That is fricken insane.
You are saying to people who backport stuff: This fixes a commit found in the following kernel releases: 6.4 6.9 3.16.68 4.4.180 4.9.137 4.14.81 4.18.19 4.19 5.11 5.7 6.6 5.10
You didn't even sort this in any sane order, how was it generated?
What in the world is anyone supposed to do with this?
If you were sent a patch with this in it, what would you think? What could you do with it?
Please be reasonable and consider us overworked stable maintainers and give us a chance to get things right. As it is, this just makes things worse...
greg k-h
Sorry, I certainly didn't want to add you more work.
I noted down which features are present in the x86_capability bitmask while they're not compiled into the kernel. Then I noted down which commits added these feature flags. So I suppose the order is from least to most significant feature bit, which now I realize doesn't help much in backporting, again sorry.
Would a more fitting Fixes: commit be the one that changed how the feature flags are used? At some point docs started stating to have them set only when features are COMPILED & HARDWARE-SUPPORTED.
What would you want to see if you had to do something with a "Fixes:" line?
thanks,
greg k-h
On 2025-07-23 at 13:57:34 +0200, Greg KH wrote:
On Wed, Jul 23, 2025 at 01:46:44PM +0200, Maciej Wieczor-Retman wrote:
On 2025-07-23 at 11:45:22 +0200, Greg KH wrote:
On Wed, Jul 23, 2025 at 11:22:49AM +0200, Maciej Wieczor-Retman wrote:
If some config options are disabled during compile time, they still are enumerated in macros that use the x86_capability bitmask - cpu_has() or this_cpu_has().
The features are also visible in /proc/cpuinfo even though they are not enabled - which is contrary to what the documentation states about the file. Examples of such feature flags are lam, fred, sgx, ibrs_enhanced, split_lock_detect, user_shstk, avx_vnni and enqcmd.
Add a DISABLED_MASK() macro that returns 32 bit chunks of the disabled feature bits bitmask.
Initialize the cpu_caps_cleared and cpu_caps_set arrays with the contents of the disabled and required bitmasks respectively. Then let apply_forced_caps() clear/set these feature bits in the x86_capability.
Fixes: 6449dcb0cac7 ("x86: CPUID and CR3/CR4 flags for Linear Address Masking") Fixes: 51c158f7aacc ("x86/cpufeatures: Add the CPU feature bit for FRED") Fixes: 706d51681d63 ("x86/speculation: Support Enhanced IBRS on future CPUs") Fixes: e7b6385b01d8 ("x86/cpufeatures: Add Intel SGX hardware bits") Fixes: 6650cdd9a8cc ("x86/split_lock: Enable split lock detection by kernel") Fixes: 701fb66d576e ("x86/cpufeatures: Add CPU feature flags for shadow stacks") Fixes: ff4f82816dff ("x86/cpufeatures: Enumerate ENQCMD and ENQCMDS instructions")
That is fricken insane.
You are saying to people who backport stuff: This fixes a commit found in the following kernel releases: 6.4 6.9 3.16.68 4.4.180 4.9.137 4.14.81 4.18.19 4.19 5.11 5.7 6.6 5.10
You didn't even sort this in any sane order, how was it generated?
What in the world is anyone supposed to do with this?
If you were sent a patch with this in it, what would you think? What could you do with it?
Please be reasonable and consider us overworked stable maintainers and give us a chance to get things right. As it is, this just makes things worse...
greg k-h
Sorry, I certainly didn't want to add you more work.
I noted down which features are present in the x86_capability bitmask while they're not compiled into the kernel. Then I noted down which commits added these feature flags. So I suppose the order is from least to most significant feature bit, which now I realize doesn't help much in backporting, again sorry.
Would a more fitting Fixes: commit be the one that changed how the feature flags are used? At some point docs started stating to have them set only when features are COMPILED & HARDWARE-SUPPORTED.
What would you want to see if you had to do something with a "Fixes:" line?
I suppose I'd want to see a Fixes:commit in a place that hasn't seen too many changes. So the backport process doesn't hit too many infrastructure changes since that makes things more tricky.
And I guess it would be great if the Fixes:commit pointed at some obvious error that happened - like a place that could dereference a NULL pointer for example.
But I thought Fixes: was supposed to mark the origin point of some error the patch is fixing?
In this case a documentation patch [1] changed how feature flags are supposed to behave. But these flags were added in various points in the past. So what should Fixes: point at then?
But anyway writing this now I get the feeling that [1] would be a better point to mark for the "Fixes:" line.
[1] ea4e3bef4c94 ("Documentation/x86: Add documentation for /proc/cpuinfo feature flags")
thanks,
greg k-h
On Wed, Jul 23, 2025 at 03:03:27PM +0200, Maciej Wieczor-Retman wrote:
On 2025-07-23 at 13:57:34 +0200, Greg KH wrote:
On Wed, Jul 23, 2025 at 01:46:44PM +0200, Maciej Wieczor-Retman wrote:
On 2025-07-23 at 11:45:22 +0200, Greg KH wrote:
On Wed, Jul 23, 2025 at 11:22:49AM +0200, Maciej Wieczor-Retman wrote:
If some config options are disabled during compile time, they still are enumerated in macros that use the x86_capability bitmask - cpu_has() or this_cpu_has().
The features are also visible in /proc/cpuinfo even though they are not enabled - which is contrary to what the documentation states about the file. Examples of such feature flags are lam, fred, sgx, ibrs_enhanced, split_lock_detect, user_shstk, avx_vnni and enqcmd.
Add a DISABLED_MASK() macro that returns 32 bit chunks of the disabled feature bits bitmask.
Initialize the cpu_caps_cleared and cpu_caps_set arrays with the contents of the disabled and required bitmasks respectively. Then let apply_forced_caps() clear/set these feature bits in the x86_capability.
Fixes: 6449dcb0cac7 ("x86: CPUID and CR3/CR4 flags for Linear Address Masking") Fixes: 51c158f7aacc ("x86/cpufeatures: Add the CPU feature bit for FRED") Fixes: 706d51681d63 ("x86/speculation: Support Enhanced IBRS on future CPUs") Fixes: e7b6385b01d8 ("x86/cpufeatures: Add Intel SGX hardware bits") Fixes: 6650cdd9a8cc ("x86/split_lock: Enable split lock detection by kernel") Fixes: 701fb66d576e ("x86/cpufeatures: Add CPU feature flags for shadow stacks") Fixes: ff4f82816dff ("x86/cpufeatures: Enumerate ENQCMD and ENQCMDS instructions")
That is fricken insane.
You are saying to people who backport stuff: This fixes a commit found in the following kernel releases: 6.4 6.9 3.16.68 4.4.180 4.9.137 4.14.81 4.18.19 4.19 5.11 5.7 6.6 5.10
You didn't even sort this in any sane order, how was it generated?
What in the world is anyone supposed to do with this?
If you were sent a patch with this in it, what would you think? What could you do with it?
Please be reasonable and consider us overworked stable maintainers and give us a chance to get things right. As it is, this just makes things worse...
greg k-h
Sorry, I certainly didn't want to add you more work.
I noted down which features are present in the x86_capability bitmask while they're not compiled into the kernel. Then I noted down which commits added these feature flags. So I suppose the order is from least to most significant feature bit, which now I realize doesn't help much in backporting, again sorry.
Would a more fitting Fixes: commit be the one that changed how the feature flags are used? At some point docs started stating to have them set only when features are COMPILED & HARDWARE-SUPPORTED.
What would you want to see if you had to do something with a "Fixes:" line?
I suppose I'd want to see a Fixes:commit in a place that hasn't seen too many changes. So the backport process doesn't hit too many infrastructure changes since that makes things more tricky.
And I guess it would be great if the Fixes:commit pointed at some obvious error that happened - like a place that could dereference a NULL pointer for example.
But I thought Fixes: was supposed to mark the origin point of some error the patch is fixing?
In this case a documentation patch [1] changed how feature flags are supposed to behave. But these flags were added in various points in the past. So what should Fixes: point at then?
But anyway writing this now I get the feeling that [1] would be a better point to mark for the "Fixes:" line.
[1] ea4e3bef4c94 ("Documentation/x86: Add documentation for /proc/cpuinfo feature flags")
If you feel that this patch should be backported to only 5.10 and newer, great, that's the correct marking place. If not, you might want to reconsider it :)
thanks,
greg k-h
On 7/23/2025 6:03 AM, Maciej Wieczor-Retman wrote:
On 2025-07-23 at 13:57:34 +0200, Greg KH wrote:
On Wed, Jul 23, 2025 at 01:46:44PM +0200, Maciej Wieczor-Retman wrote:
On 2025-07-23 at 11:45:22 +0200, Greg KH wrote:
On Wed, Jul 23, 2025 at 11:22:49AM +0200, Maciej Wieczor-Retman wrote:
If some config options are disabled during compile time, they still are enumerated in macros that use the x86_capability bitmask - cpu_has() or this_cpu_has().
The features are also visible in /proc/cpuinfo even though they are not enabled - which is contrary to what the documentation states about the file. Examples of such feature flags are lam, fred, sgx, ibrs_enhanced, split_lock_detect, user_shstk, avx_vnni and enqcmd.
Add a DISABLED_MASK() macro that returns 32 bit chunks of the disabled feature bits bitmask.
Initialize the cpu_caps_cleared and cpu_caps_set arrays with the contents of the disabled and required bitmasks respectively. Then let apply_forced_caps() clear/set these feature bits in the x86_capability.
Fixes: 6449dcb0cac7 ("x86: CPUID and CR3/CR4 flags for Linear Address Masking") Fixes: 51c158f7aacc ("x86/cpufeatures: Add the CPU feature bit for FRED") Fixes: 706d51681d63 ("x86/speculation: Support Enhanced IBRS on future CPUs") Fixes: e7b6385b01d8 ("x86/cpufeatures: Add Intel SGX hardware bits") Fixes: 6650cdd9a8cc ("x86/split_lock: Enable split lock detection by kernel") Fixes: 701fb66d576e ("x86/cpufeatures: Add CPU feature flags for shadow stacks") Fixes: ff4f82816dff ("x86/cpufeatures: Enumerate ENQCMD and ENQCMDS instructions")
That is fricken insane.
You are saying to people who backport stuff: This fixes a commit found in the following kernel releases: 6.4 6.9 3.16.68 4.4.180 4.9.137 4.14.81 4.18.19 4.19 5.11 5.7 6.6 5.10
You didn't even sort this in any sane order, how was it generated?
What in the world is anyone supposed to do with this?
If you were sent a patch with this in it, what would you think? What could you do with it?
Please be reasonable and consider us overworked stable maintainers and give us a chance to get things right. As it is, this just makes things worse...
greg k-h
Sorry, I certainly didn't want to add you more work.
I noted down which features are present in the x86_capability bitmask while they're not compiled into the kernel. Then I noted down which commits added these feature flags. So I suppose the order is from least to most significant feature bit, which now I realize doesn't help much in backporting, again sorry.
Would a more fitting Fixes: commit be the one that changed how the feature flags are used? At some point docs started stating to have them set only when features are COMPILED & HARDWARE-SUPPORTED.
What would you want to see if you had to do something with a "Fixes:" line?
I suppose I'd want to see a Fixes:commit in a place that hasn't seen too many changes. So the backport process doesn't hit too many infrastructure changes since that makes things more tricky.
And I guess it would be great if the Fixes:commit pointed at some obvious error that happened - like a place that could dereference a NULL pointer for example.
But I thought Fixes: was supposed to mark the origin point of some error the patch is fixing?
In this case a documentation patch [1] changed how feature flags are supposed to behave. But these flags were added in various points in the past. So what should Fixes: point at then?
But anyway writing this now I get the feeling that [1] would be a better point to mark for the "Fixes:" line.
[1] ea4e3bef4c94 ("Documentation/x86: Add documentation for /proc/cpuinfo feature flags")
We need to investigate the root cause of this bug; and this seems like a regression caused by my patch set that automatically generates CPU feature masks.
Just further debugging is needed to identify the culprit.
Thanks! Xin
On 7/23/2025 8:52 AM, Xin Li wrote:
We need to investigate the root cause of this bug; and this seems like a regression caused by my patch set that automatically generates CPU feature masks.
My bad, it's not caused by my patch set.
The bug is there for a long time before the patch set was merged.
And it seems we need a different patch for older Linux kernels.
Thanks! Xin
On Wed, Jul 23, 2025 at 11:22:49AM +0200, Maciej Wieczor-Retman wrote:
+static __init void init_cpu_cap(struct cpuinfo_x86 *c) +{
- int i;
- for (i = 0; i < NCAPINTS; i++) {
cpu_caps_set[i] = REQUIRED_MASK(i);cpu_caps_cleared[i] = DISABLED_MASK(i);- }
+}
There's already apply_forced_caps(). Not another cap massaging function please. Add that stuff there.
As to what the Fixes: tag should be - it should not have any Fixes: tag because AFAICT, this has always been this way. So this fix should be backported everywhere.
Thx.
On 2025-07-23 at 15:46:40 +0200, Borislav Petkov wrote:
On Wed, Jul 23, 2025 at 11:22:49AM +0200, Maciej Wieczor-Retman wrote:
+static __init void init_cpu_cap(struct cpuinfo_x86 *c) +{
- int i;
- for (i = 0; i < NCAPINTS; i++) {
cpu_caps_set[i] = REQUIRED_MASK(i);cpu_caps_cleared[i] = DISABLED_MASK(i);- }
+}
There's already apply_forced_caps(). Not another cap massaging function please. Add that stuff there.
I'll try that, but can't it overwrite some things? apply_forced_caps() is called three times and cpu_caps_set/cleared are modified in between from what I can see. init_cpu_cap() was supposed to only initialize these arrays.
As to what the Fixes: tag should be - it should not have any Fixes: tag because AFAICT, this has always been this way. So this fix should be backported everywhere.
I found that in 5.9-rc1 the documentation for how /proc/cpuinfo should work was merged [1]. I understand that from that point on, while one can't rely on a feature's absence, it's a reliable convention that if a flag is present, then the feature is working. So from 5.9 on, it seems like a bug when these features show up as working while they're not due to not being compiled.
[1] ea4e3bef4c94 ("Documentation/x86: Add documentation for /proc/cpuinfo feature flags")
Thx.
-- Regards/Gruss, Boris.
On July 23, 2025 8:13:07 AM PDT, Maciej Wieczor-Retman maciej.wieczor-retman@intel.com wrote:
On 2025-07-23 at 15:46:40 +0200, Borislav Petkov wrote:
On Wed, Jul 23, 2025 at 11:22:49AM +0200, Maciej Wieczor-Retman wrote:
+static __init void init_cpu_cap(struct cpuinfo_x86 *c) +{
- int i;
- for (i = 0; i < NCAPINTS; i++) {
cpu_caps_set[i] = REQUIRED_MASK(i);cpu_caps_cleared[i] = DISABLED_MASK(i);- }
+}
There's already apply_forced_caps(). Not another cap massaging function please. Add that stuff there.
I'll try that, but can't it overwrite some things? apply_forced_caps() is called three times and cpu_caps_set/cleared are modified in between from what I can see. init_cpu_cap() was supposed to only initialize these arrays.
As to what the Fixes: tag should be - it should not have any Fixes: tag because AFAICT, this has always been this way. So this fix should be backported everywhere.
I found that in 5.9-rc1 the documentation for how /proc/cpuinfo should work was merged [1]. I understand that from that point on, while one can't rely on a feature's absence, it's a reliable convention that if a flag is present, then the feature is working. So from 5.9 on, it seems like a bug when these features show up as working while they're not due to not being compiled.
[1] ea4e3bef4c94 ("Documentation/x86: Add documentation for /proc/cpuinfo feature flags")
Thx.
-- Regards/Gruss, Boris.
What are you concerned it would overwrite? I'm confused.
On 2025-07-23 at 08:28:32 -0700, H. Peter Anvin wrote:
On July 23, 2025 8:13:07 AM PDT, Maciej Wieczor-Retman maciej.wieczor-retman@intel.com wrote:
On 2025-07-23 at 15:46:40 +0200, Borislav Petkov wrote:
On Wed, Jul 23, 2025 at 11:22:49AM +0200, Maciej Wieczor-Retman wrote:
+static __init void init_cpu_cap(struct cpuinfo_x86 *c) +{
- int i;
- for (i = 0; i < NCAPINTS; i++) {
cpu_caps_set[i] = REQUIRED_MASK(i);cpu_caps_cleared[i] = DISABLED_MASK(i);- }
+}
There's already apply_forced_caps(). Not another cap massaging function please. Add that stuff there.
I'll try that, but can't it overwrite some things? apply_forced_caps() is called three times and cpu_caps_set/cleared are modified in between from what I can see. init_cpu_cap() was supposed to only initialize these arrays.
What are you concerned it would overwrite? I'm confused.
I thought that cpu_caps_set/cleared could change in-between apply_forced_caps() calls. Therefore if we also applied the DISABLED_MASK() in every apply_forced_caps() call I thought it might clear some flag that other function might set.
But I've been looking at these calls for a while now and that doesn't seem possible. Changes are made only if features are compiled, so it doesn't interfere with the DISABLED_MASK().
Sorry for the confusion.
On July 23, 2025 10:13:52 AM PDT, Maciej Wieczor-Retman maciej.wieczor-retman@intel.com wrote:
On 2025-07-23 at 08:28:32 -0700, H. Peter Anvin wrote:
On July 23, 2025 8:13:07 AM PDT, Maciej Wieczor-Retman maciej.wieczor-retman@intel.com wrote:
On 2025-07-23 at 15:46:40 +0200, Borislav Petkov wrote:
On Wed, Jul 23, 2025 at 11:22:49AM +0200, Maciej Wieczor-Retman wrote:
+static __init void init_cpu_cap(struct cpuinfo_x86 *c) +{
- int i;
- for (i = 0; i < NCAPINTS; i++) {
cpu_caps_set[i] = REQUIRED_MASK(i);cpu_caps_cleared[i] = DISABLED_MASK(i);- }
+}
There's already apply_forced_caps(). Not another cap massaging function please. Add that stuff there.
I'll try that, but can't it overwrite some things? apply_forced_caps() is called three times and cpu_caps_set/cleared are modified in between from what I can see. init_cpu_cap() was supposed to only initialize these arrays.
What are you concerned it would overwrite? I'm confused.
I thought that cpu_caps_set/cleared could change in-between apply_forced_caps() calls. Therefore if we also applied the DISABLED_MASK() in every apply_forced_caps() call I thought it might clear some flag that other function might set.
But I've been looking at these calls for a while now and that doesn't seem possible. Changes are made only if features are compiled, so it doesn't interfere with the DISABLED_MASK().
Sorry for the confusion.
Any changes would be additive, or we would be in a world of hurt anyway.
On 2025-07-23 02:22, Maciej Wieczor-Retman wrote:
arch/x86/kernel/cpu/common.c | 12 ++++++++++++ arch/x86/tools/cpufeaturemasks.awk | 8 ++++++++ 2 files changed, 20 insertions(+)
diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c index 77afca95cced..ba8b5fba8552 100644 --- a/arch/x86/kernel/cpu/common.c +++ b/arch/x86/kernel/cpu/common.c @@ -1709,6 +1709,16 @@ static void __init cpu_parse_early_param(void) } } +static __init void init_cpu_cap(struct cpuinfo_x86 *c) +{
- int i;
- for (i = 0; i < NCAPINTS; i++) {
cpu_caps_set[i] = REQUIRED_MASK(i);cpu_caps_cleared[i] = DISABLED_MASK(i);- }
+}
No... just use an static array initializer for cpu_caps_cleared[]. You don't even need to add any code at all.
Ironically enough, I actually had those macros generated in the original awk script version, but they weren't used.
And you MUST NOT initialize cpu_caps_set[]. What we *can* and probably should do is, at the very end of the process, verify that the final mask conforms to cpu_caps_set[] and print a nasty message and set the TAINT_CPU_OUT_OF_SPEC flag:
add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK)
... but that is a separate patch, and doesn't need to be backported.
Xin send me a patch privately (attached) but it needs to have the REQUIRED_MASK_INIT_VALUES part removed and made into a proper patch.
Xin didn't add an SoB on it, but you can use mine:
Signed-off-by: H. Peter Anvin (Intel) hpa@zytor.com
-hpa
On 2025-07-23 at 07:08:43 -0700, H. Peter Anvin wrote:
On 2025-07-23 02:22, Maciej Wieczor-Retman wrote:
arch/x86/kernel/cpu/common.c | 12 ++++++++++++ arch/x86/tools/cpufeaturemasks.awk | 8 ++++++++ 2 files changed, 20 insertions(+)
diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c index 77afca95cced..ba8b5fba8552 100644 --- a/arch/x86/kernel/cpu/common.c +++ b/arch/x86/kernel/cpu/common.c @@ -1709,6 +1709,16 @@ static void __init cpu_parse_early_param(void) } } +static __init void init_cpu_cap(struct cpuinfo_x86 *c) +{
- int i;
- for (i = 0; i < NCAPINTS; i++) {
cpu_caps_set[i] = REQUIRED_MASK(i);cpu_caps_cleared[i] = DISABLED_MASK(i);- }
+}
No... just use an static array initializer for cpu_caps_cleared[]. You don't even need to add any code at all.
Oh, right, that seems simpler
Ironically enough, I actually had those macros generated in the original awk script version, but they weren't used.
And you MUST NOT initialize cpu_caps_set[].
After thinking about it more I can see why doing that could be a problem, yes.
What we *can* and probably should do is, at the very end of the process, verify that the final mask conforms to cpu_caps_set[] and print a nasty message and set the TAINT_CPU_OUT_OF_SPEC flag:
add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK)
Where would that end be though? End of identify_cpu()? apply_forced_caps() is almost at the end of the function, and other functions don't seem to modify x86_cpuinfo or cpu_caps_set[].
... but that is a separate patch, and doesn't need to be backported.
Xin send me a patch privately (attached) but it needs to have the REQUIRED_MASK_INIT_VALUES part removed and made into a proper patch.
Xin didn't add an SoB on it, but you can use mine:
Signed-off-by: H. Peter Anvin (Intel) hpa@zytor.com
Thanks, I'll work with that.
-hpa
linux-stable-mirror@lists.linaro.org
-
Borislav Petkov -
Greg KH -
H. Peter Anvin -
Maciej Wieczor-Retman -
Xin Li