Linux 5.16 compat: don't use XSTATE_XSAVE to save FPU state #13059
Conversation
|
Sorry, pushed from wrong branch, fixed now. If there is a need to handle exceptions as the previous version was doing, I can try to wrap my head around the |
|
Might it make sense to switch over unconditionally, not just for 5.16+, if there's no important downsides? (Or, if we think there might be but we haven't spotted them yet, to rip off the band-aid and spot them faster?) |
|
Yes, it might, I just choose the least intrusive approach to play it safe. There may also be a slight performance loss involved since the |
| @@ -322,7 +441,7 @@ kfpu_end(void) | |||
| #define kfpu_init() 0 | |||
| #define kfpu_fini() ((void) 0) | |||
|
|
|||
| #endif /* defined(HAVE_KERNEL_FPU_INTERNAL) */ | |||
| #endif /* defined HAVE_KERNEL_FPU_INTERNAL || HAVE_KERNEL_FPU_XSAVE INTERNAL */ | |||
There was a problem hiding this comment.
space in HAVE_KERNEL_FPU_XSAVE INTERNAL (also a better phrasing would be defined(HAVE_KERNEL_FPU_INTERNAL) || defined(HAVE_KERNEL_FPU_XSAVE_INTERNAL) or defined(HAVE_KERNEL_FPU_INTERNAL || HAVE_KERNEL_FPU_XSAVE_INTERNAL))
There was a problem hiding this comment.
Thanks for catching that, fixed.
I would vote for this as well. That way it gets exercised by buildbot for all kernels and simplifies the code. If there is a slight performance loss, that's fine, since it's just something we're going to have to live with going forward anyway. |
|
Applying these changes to just 5.16+ kernels would be the lowest risk option, and it's consistent with how we've tried to handle this in the past. But in this case I could see making an exception. Applying this to all kernels which don't provide fpu symbols looks like it would significantly simplify the code which would be nice. Perhaps a reasonable compromise would be to apply this patch to master which only using the new code for 5.16+. This could then be easily backported for the 2.1.3 release. Then follow up with a commit to master which uses the new code unconditionally. Which is really what we'd prefer to maintain long term. Previously we found that running the I also agree for the reasons you mentioned it's reasonable to not handle exceptions here. Thanks again for tackling this! |
|
Yes, I think this is a good idea, that way we would make sure that changes to stable releases are minimal, while having a broader range of systems running master exercising the new code. Thinking more of it, I could come up with an edge case where things may break: a very old kernel with eagerfpu=off will use the CR0.TS bit during context switching and calling XSAVE with CR0.TS=1 will generate a #NM fault. So if this kernel happens to have backported the GPL only kfpu patch, it will panic. Not sure if we support such a setup at all. It's a very unlikely scenario also, and no one would intentionally set eagerfpu=off on a CPU supporting SIMD extension since manipulating the CT0.TS bits takes longer than storing the FPU state. So not sure if we should add a test for Setting up some test datasets with encryption, fletcher, sha2, skein and edonr checksums on a raidz pool and doing I/O in loop on them while running mprime is on my todo list. That should cover all code paths which use |
|
If you get a chance to rebase and force update this PR this should clear up the CI failures, |
|
I added a regression test to make sure FPU support won't break without notice again, not sure where to put it though. I'd appreciate any hint. Once I know where to put the test I'll squash and rebase. |
|
The tests didn't went so well, I'm seeing system freezes. Have to look into it more deeply. |
|
D'oh once you push from a wrong branch things will start to go south.. Will run the tests with the proper version tonight. |
|
Depending on exactly what the test case looks like I could see it/them living under |
|
It simply test if it's running under Linux on a x86 box with SSE support. If so, it checks if As a side note, shouldn't |
|
The |
All right, will put it there.
I see, thanks for the explanation. |
|
I forget if I mentioned in this thread, but my intent after this gets fixed
was to make a configure flag that would make the FPU configure checks hard
fail on "not found' instead of continuing - that way, we can catch those on
the build-only bots, too.
Not suggesting the two are mutually exclusive, just mentioning as a data
point for where I was planning to catch this.
…On Mon, Feb 7, 2022 at 2:38 PM Attila Fülöp ***@***.***> wrote:
functional/simd/ sounds good.
All right, will put it there.
The Requires-builders: style comment is currently only understood by the
buildbit CI builders, and not the GitHub actions CI builders. So those
always run.
I see, thanks for the explanation.
—
Reply to this email directly, view it on GitHub
<#13059 (comment)>, or
unsubscribe
<https://github.com/notifications/unsubscribe-auth/AABUI7LDV6W3Z5LBSPEYTILU2ANTNANCNFSM5NOOF47A>
.
Triage notifications on the go with GitHub Mobile for iOS
<https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675>
or Android
<https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.
You are receiving this because you commented.Message ID:
***@***.***>
|
|
Not that I know of. That may make sense, just trying to think of some edge cases where this may come in the way. Maybe an embedded system with a striped down Linux kernel? Does this make sense? I think if you make it hard fail, a configure option like |
|
Oh of course, I wasn't suggesting you not add something already written and
fit for purpose, and I'm sorry if it came off that way.
…On Mon, Feb 7, 2022 at 4:28 PM Attila Fülöp ***@***.***> wrote:
Not that I know of. That may make sense, just trying to think of some edge
cases where this may come in the way. Maybe an embedded system with a
striped down Linux kernel? Does this make sense? I think if you make it
hard fail, a configure option like --without-fpu should be put in place.
Of course that would collide with a zfs-test. Since I wrote the test
already, let me add it. If need be, we can remove it later with no fuss.
—
Reply to this email directly, view it on GitHub
<#13059 (comment)>, or
unsubscribe
<https://github.com/notifications/unsubscribe-auth/AABUI7LSWR76HGIRZ7OK4TTU2A2QRANCNFSM5NOOF47A>
.
Triage notifications on the go with GitHub Mobile for iOS
<https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675>
or Android
<https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.
You are receiving this because you commented.Message ID:
***@***.***>
|
|
No sorry, you didn't come off that way. What I wanted to say was that it's a good idea, and you shouldn't feel that the zfs-test is in your way (and just go ahead as you see fit). I've to leave the room now, four fio and a mprime torture run make for really noisy box :-) |
AttilaFueloep
force-pushed
the
zfs-add-xsave
branch
from
0b9efb7 to
6de32e2
Compare
|
Squashed and rebased. Added a regression test. |
|
Tested this PR now. Made a 3 disk (actually 3 partition on the same SSD) raidz1 pool, created three datasets on it, one encrypted/fletcher3, one unencrypted/sha512 and one unencrypted/skein. Ran a 4 thread/16 files All went well, no |
|
This is ready for review, looks like I can't remove the draft tag. |
|
IIRC there should be a button around the CI status window at the bottom for "ready for review? mark as not draft any more" - it has a notice that only people with write access to the branch can make that change, for me, here. |
|
D'oh, yes there is, thanks! |
|
@AttilaFueloep thanks for tackling this one, merged. |
|
As discussed here, I'll open a new PR, factoring out the XSTATE stuff for all kernels not exporting *kfpu_{begin,end}.
I think the above mentioned PR will be fail-safe per se by not using any kernel fpu stuff. |
Cleanup the kernel SIMD code by removing kernel dependencies.
- Replace XSTATE_XSAVE with our own XSAVE implementation for all
kernels not exporting kernel_fpu{begin,end}(), see openzfs#13059
- Replace union fpregs_state by a uint8_t * buffer and get the size
of the buffer from the hardware via the CPUID instruction
- Replace kernels xgetbv() by our own implementation which was
already there for userspace.
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Linux 5.16 moved XSTATE_XSAVE and XSTATE_XRESTORE out of our reach,
so add our own XSAVE{,OPT,S} code and use it for Linux 5.16.
Please note that this differs from previous behavior in that it
won't handle exceptions created by XSAVE an XRSTOR. This is sensible
for three reasons.
- Exceptions during XSAVE and XRSTOR can only occur if the feature
is not supported or enabled or the memory operand isn't aligned
on a 64 byte boundary. If this happens something else went
terribly wrong, and it may be better to stop execution.
- Previously we just printed a warning and didn't handle the fault,
this is arguable for the above reason.
- All other *SAVE instruction also don't handle exceptions, so this
at least aligns behavior.
Finally add a test to catch such a regression in the future.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13042
Closes openzfs#13059
Linux 5.16 moved XSTATE_XSAVE and XSTATE_XRESTORE out of our reach,
so add our own XSAVE{,OPT,S} code and use it for Linux 5.16.
Please note that this differs from previous behavior in that it
won't handle exceptions created by XSAVE an XRSTOR. This is sensible
for three reasons.
- Exceptions during XSAVE and XRSTOR can only occur if the feature
is not supported or enabled or the memory operand isn't aligned
on a 64 byte boundary. If this happens something else went
terribly wrong, and it may be better to stop execution.
- Previously we just printed a warning and didn't handle the fault,
this is arguable for the above reason.
- All other *SAVE instruction also don't handle exceptions, so this
at least aligns behavior.
Finally add a test to catch such a regression in the future.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13042
Closes openzfs#13059
Linux 5.16 moved XSTATE_XSAVE and XSTATE_XRESTORE out of our reach,
so add our own XSAVE{,OPT,S} code and use it for Linux 5.16.
Please note that this differs from previous behavior in that it
won't handle exceptions created by XSAVE an XRSTOR. This is sensible
for three reasons.
- Exceptions during XSAVE and XRSTOR can only occur if the feature
is not supported or enabled or the memory operand isn't aligned
on a 64 byte boundary. If this happens something else went
terribly wrong, and it may be better to stop execution.
- Previously we just printed a warning and didn't handle the fault,
this is arguable for the above reason.
- All other *SAVE instruction also don't handle exceptions, so this
at least aligns behavior.
Finally add a test to catch such a regression in the future.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13042
Closes openzfs#13059
Cleanup the kernel SIMD code by removing kernel dependencies.
- Replace XSTATE_XSAVE with our own XSAVE implementation for all
kernels not exporting kernel_fpu{begin,end}(), see openzfs#13059
- Replace union fpregs_state by a uint8_t * buffer and get the size
of the buffer from the hardware via the CPUID instruction
- Replace kernels xgetbv() by our own implementation which was
already there for userspace.
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Linux 5.16 moved XSTATE_XSAVE and XSTATE_XRESTORE out of our reach,
so add our own XSAVE{,OPT,S} code and use it for Linux 5.16.
Please note that this differs from previous behavior in that it
won't handle exceptions created by XSAVE an XRSTOR. This is sensible
for three reasons.
- Exceptions during XSAVE and XRSTOR can only occur if the feature
is not supported or enabled or the memory operand isn't aligned
on a 64 byte boundary. If this happens something else went
terribly wrong, and it may be better to stop execution.
- Previously we just printed a warning and didn't handle the fault,
this is arguable for the above reason.
- All other *SAVE instruction also don't handle exceptions, so this
at least aligns behavior.
Finally add a test to catch such a regression in the future.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13042
Closes openzfs#13059
Cleanup the kernel SIMD code by removing kernel dependencies.
- Replace XSTATE_XSAVE with our own XSAVE implementation for all
kernels not exporting kernel_fpu{begin,end}(), see openzfs#13059
- Replace union fpregs_state by a uint8_t * buffer and get the size
of the buffer from the hardware via the CPUID instruction
- Replace kernels xgetbv() by our own implementation which was
already there for userspace.
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Cleanup the kernel SIMD code by removing kernel dependencies.
- Replace XSTATE_XSAVE with our own XSAVE implementation for all
kernels not exporting kernel_fpu{begin,end}(), see openzfs#13059
- Replace union fpregs_state by a uint8_t * buffer and get the size
of the buffer from the hardware via the CPUID instruction
- Replace kernels xgetbv() by our own implementation which was
already there for userspace.
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Cleanup the kernel SIMD code by removing kernel dependencies.
- Replace XSTATE_XSAVE with our own XSAVE implementation for all
kernels not exporting kernel_fpu{begin,end}(), see #13059
- Replace union fpregs_state by a uint8_t * buffer and get the size
of the buffer from the hardware via the CPUID instruction
- Replace kernels xgetbv() by our own implementation which was
already there for userspace.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes #13102
Linux 5.16 moved XSTATE_XSAVE and XSTATE_XRESTORE out of our reach,
so add our own XSAVE{,OPT,S} code and use it for Linux 5.16.
Please note that this differs from previous behavior in that it
won't handle exceptions created by XSAVE an XRSTOR. This is sensible
for three reasons.
- Exceptions during XSAVE and XRSTOR can only occur if the feature
is not supported or enabled or the memory operand isn't aligned
on a 64 byte boundary. If this happens something else went
terribly wrong, and it may be better to stop execution.
- Previously we just printed a warning and didn't handle the fault,
this is arguable for the above reason.
- All other *SAVE instruction also don't handle exceptions, so this
at least aligns behavior.
Finally add a test to catch such a regression in the future.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13042
Closes openzfs#13059
Cleanup the kernel SIMD code by removing kernel dependencies.
- Replace XSTATE_XSAVE with our own XSAVE implementation for all
kernels not exporting kernel_fpu{begin,end}(), see openzfs#13059
- Replace union fpregs_state by a uint8_t * buffer and get the size
of the buffer from the hardware via the CPUID instruction
- Replace kernels xgetbv() by our own implementation which was
already there for userspace.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13102
Linux 5.16 moved XSTATE_XSAVE and XSTATE_XRESTORE out of our reach,
so add our own XSAVE{,OPT,S} code and use it for Linux 5.16.
Please note that this differs from previous behavior in that it
won't handle exceptions created by XSAVE an XRSTOR. This is sensible
for three reasons.
- Exceptions during XSAVE and XRSTOR can only occur if the feature
is not supported or enabled or the memory operand isn't aligned
on a 64 byte boundary. If this happens something else went
terribly wrong, and it may be better to stop execution.
- Previously we just printed a warning and didn't handle the fault,
this is arguable for the above reason.
- All other *SAVE instruction also don't handle exceptions, so this
at least aligns behavior.
Finally add a test to catch such a regression in the future.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13042
Closes openzfs#13059
Cleanup the kernel SIMD code by removing kernel dependencies.
- Replace XSTATE_XSAVE with our own XSAVE implementation for all
kernels not exporting kernel_fpu{begin,end}(), see openzfs#13059
- Replace union fpregs_state by a uint8_t * buffer and get the size
of the buffer from the hardware via the CPUID instruction
- Replace kernels xgetbv() by our own implementation which was
already there for userspace.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13102
Cleanup the kernel SIMD code by removing kernel dependencies.
- Replace XSTATE_XSAVE with our own XSAVE implementation for all
kernels not exporting kernel_fpu{begin,end}(), see openzfs#13059
- Replace union fpregs_state by a uint8_t * buffer and get the size
of the buffer from the hardware via the CPUID instruction
- Replace kernels xgetbv() by our own implementation which was
already there for userspace.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13102
Cleanup the kernel SIMD code by removing kernel dependencies.
- Replace XSTATE_XSAVE with our own XSAVE implementation for all
kernels not exporting kernel_fpu{begin,end}(), see openzfs#13059
- Replace union fpregs_state by a uint8_t * buffer and get the size
of the buffer from the hardware via the CPUID instruction
- Replace kernels xgetbv() by our own implementation which was
already there for userspace.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13102
Cleanup the kernel SIMD code by removing kernel dependencies.
- Replace XSTATE_XSAVE with our own XSAVE implementation for all
kernels not exporting kernel_fpu{begin,end}(), see openzfs#13059
- Replace union fpregs_state by a uint8_t * buffer and get the size
of the buffer from the hardware via the CPUID instruction
- Replace kernels xgetbv() by our own implementation which was
already there for userspace.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13102
Cleanup the kernel SIMD code by removing kernel dependencies.
- Replace XSTATE_XSAVE with our own XSAVE implementation for all
kernels not exporting kernel_fpu{begin,end}(), see openzfs#13059
- Replace union fpregs_state by a uint8_t * buffer and get the size
of the buffer from the hardware via the CPUID instruction
- Replace kernels xgetbv() by our own implementation which was
already there for userspace.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13102
Cleanup the kernel SIMD code by removing kernel dependencies.
- Replace XSTATE_XSAVE with our own XSAVE implementation for all
kernels not exporting kernel_fpu{begin,end}(), see openzfs#13059
- Replace union fpregs_state by a uint8_t * buffer and get the size
of the buffer from the hardware via the CPUID instruction
- Replace kernels xgetbv() by our own implementation which was
already there for userspace.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13102
Cleanup the kernel SIMD code by removing kernel dependencies.
- Replace XSTATE_XSAVE with our own XSAVE implementation for all
kernels not exporting kernel_fpu{begin,end}(), see openzfs#13059
- Replace union fpregs_state by a uint8_t * buffer and get the size
of the buffer from the hardware via the CPUID instruction
- Replace kernels xgetbv() by our own implementation which was
already there for userspace.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13102
Cleanup the kernel SIMD code by removing kernel dependencies.
- Replace XSTATE_XSAVE with our own XSAVE implementation for all
kernels not exporting kernel_fpu{begin,end}(), see openzfs#13059
- Replace union fpregs_state by a uint8_t * buffer and get the size
of the buffer from the hardware via the CPUID instruction
- Replace kernels xgetbv() by our own implementation which was
already there for userspace.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13102
Cleanup the kernel SIMD code by removing kernel dependencies.
- Replace XSTATE_XSAVE with our own XSAVE implementation for all
kernels not exporting kernel_fpu{begin,end}(), see openzfs#13059
- Replace union fpregs_state by a uint8_t * buffer and get the size
of the buffer from the hardware via the CPUID instruction
- Replace kernels xgetbv() by our own implementation which was
already there for userspace.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13102
Cleanup the kernel SIMD code by removing kernel dependencies.
- Replace XSTATE_XSAVE with our own XSAVE implementation for all
kernels not exporting kernel_fpu{begin,end}(), see openzfs#13059
- Replace union fpregs_state by a uint8_t * buffer and get the size
of the buffer from the hardware via the CPUID instruction
- Replace kernels xgetbv() by our own implementation which was
already there for userspace.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13102
Cleanup the kernel SIMD code by removing kernel dependencies.
- Replace XSTATE_XSAVE with our own XSAVE implementation for all
kernels not exporting kernel_fpu{begin,end}(), see openzfs#13059
- Replace union fpregs_state by a uint8_t * buffer and get the size
of the buffer from the hardware via the CPUID instruction
- Replace kernels xgetbv() by our own implementation which was
already there for userspace.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13102
Cleanup the kernel SIMD code by removing kernel dependencies.
- Replace XSTATE_XSAVE with our own XSAVE implementation for all
kernels not exporting kernel_fpu{begin,end}(), see openzfs#13059
- Replace union fpregs_state by a uint8_t * buffer and get the size
of the buffer from the hardware via the CPUID instruction
- Replace kernels xgetbv() by our own implementation which was
already there for userspace.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13102
Cleanup the kernel SIMD code by removing kernel dependencies.
- Replace XSTATE_XSAVE with our own XSAVE implementation for all
kernels not exporting kernel_fpu{begin,end}(), see openzfs#13059
- Replace union fpregs_state by a uint8_t * buffer and get the size
of the buffer from the hardware via the CPUID instruction
- Replace kernels xgetbv() by our own implementation which was
already there for userspace.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13102
Cleanup the kernel SIMD code by removing kernel dependencies.
- Replace XSTATE_XSAVE with our own XSAVE implementation for all
kernels not exporting kernel_fpu{begin,end}(), see openzfs#13059
- Replace union fpregs_state by a uint8_t * buffer and get the size
of the buffer from the hardware via the CPUID instruction
- Replace kernels xgetbv() by our own implementation which was
already there for userspace.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes openzfs#13102
Motivation and Context
Linux 5.16 moved XSTATE_XSAVE and XSTATE_XRESTORE out of our reach.
Closes #13042
Description
Add our own XSAVE{,OPT,S} code and use it for Linux 5.16.
Please note that this differs from previous behavior in that it
won't handle exception created by XSAVE an XRSTOR. This is sensible
for three reasons.
Exceptions during XSAVE and XRSTOR can only occur if the feature
is not supported or enabled or the memory operand isn't aligned
on a 64 byte boundary. If this happens something else went
terribly wrong, and it may be better to stop execution.
Previously we just printed a warning and didn't handle the fault,
this is arguable for the above reason.
All other *SAVE instruction also don't handle exceptions, so this
at least aligns behavior.
How Has This Been Tested?
Verified with gdb that the FPU registers are properly restored.
Types of changes
Checklist:
Signed-off-by.