| Message ID | 20230127174552.3370169-1-nsg@linux.ibm.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [RFC,v1] KVM: selftests: Compile s390 tests with -march=z10 | expand |
On 27/01/2023 18.45, Nina Schoetterl-Glausch wrote: > The guest used in s390 kvm selftests is not be set up to handle all > instructions the compiler might emit, i.e. vector instructions, leading > to crashes. > Limit what the compiler emits to the oldest machine model currently > supported by Linux. > > Signed-off-by: Nina Schoetterl-Glausch <nsg@linux.ibm.com> > --- > > Should we also set -mtune? I don't think it's really necessary > Since it are vector instructions that caused the problem here, there > are some alternatives: > * use -mno-vx > * set the required guest control bit to enable vector instructions on > models supporting them > > -march=z10 might prevent similar issues with other instructions, but I > don't know if there actually exist other relevant instructions, so it > could be needlessly restricting. FWIW, the vector instructions have been introduced with the z13 ... so limiting to the zEC12 could be enough. > diff --git a/tools/testing/selftests/kvm/Makefile b/tools/testing/selftests/kvm/Makefile > index 1750f91dd936..df0989949eb5 100644 > --- a/tools/testing/selftests/kvm/Makefile > +++ b/tools/testing/selftests/kvm/Makefile > @@ -200,6 +200,9 @@ CFLAGS += -Wall -Wstrict-prototypes -Wuninitialized -O2 -g -std=gnu99 \ > -I$(LINUX_TOOL_ARCH_INCLUDE) -I$(LINUX_HDR_PATH) -Iinclude \ > -I$(<D) -Iinclude/$(ARCH_DIR) -I ../rseq -I.. $(EXTRA_CFLAGS) \ > $(KHDR_INCLUDES) > +ifeq ($(ARCH),s390) > + CFLAGS += -march=z10 > +endif Starting with z10 sounds sane to me, we still can adjust later if necessary. Reviewed-by: Thomas Huth <thuth@redhat.com>
On Fri, 2023-01-27 at 19:12 +0100, Thomas Huth wrote: > On 27/01/2023 18.45, Nina Schoetterl-Glausch wrote: > > The guest used in s390 kvm selftests is not be set up to handle all > > instructions the compiler might emit, i.e. vector instructions, leading > > to crashes. > > Limit what the compiler emits to the oldest machine model currently > > supported by Linux. > > > > Signed-off-by: Nina Schoetterl-Glausch <nsg@linux.ibm.com> > > --- > > > > Should we also set -mtune? > > I don't think it's really necessary > > > Since it are vector instructions that caused the problem here, there > > are some alternatives: > > * use -mno-vx > > * set the required guest control bit to enable vector instructions on > > models supporting them > > > > -march=z10 might prevent similar issues with other instructions, but I > > don't know if there actually exist other relevant instructions, so it > > could be needlessly restricting. > > FWIW, the vector instructions have been introduced with the z13 ... so > limiting to the zEC12 could be enough. Yes, however, if we only want to fix the issue with regards to vector instructions, one of the alternatives would be a better solution, IMO. With regards to the second, I'm not sure what happens on old models if we unconditionally enable the control bit, I'm guessing it'd be fine. The question is how likely similar issues are and if we're fine with possibly running into them. For what it's worth, finding the cause here was easy, tracing kvm-s390 events showed the faulting instruction. > > > diff --git a/tools/testing/selftests/kvm/Makefile b/tools/testing/selftests/kvm/Makefile > > index 1750f91dd936..df0989949eb5 100644 > > --- a/tools/testing/selftests/kvm/Makefile > > +++ b/tools/testing/selftests/kvm/Makefile > > @@ -200,6 +200,9 @@ CFLAGS += -Wall -Wstrict-prototypes -Wuninitialized -O2 -g -std=gnu99 \ > > -I$(LINUX_TOOL_ARCH_INCLUDE) -I$(LINUX_HDR_PATH) -Iinclude \ > > -I$(<D) -Iinclude/$(ARCH_DIR) -I ../rseq -I.. $(EXTRA_CFLAGS) \ > > $(KHDR_INCLUDES) > > +ifeq ($(ARCH),s390) > > + CFLAGS += -march=z10 > > +endif > > Starting with z10 sounds sane to me, we still can adjust later if necessary. > > Reviewed-by: Thomas Huth <thuth@redhat.com> > Thanks!
On 1/27/23 18:45, Nina Schoetterl-Glausch wrote: > The guest used in s390 kvm selftests is not be set up to handle all > instructions the compiler might emit, i.e. vector instructions, leading > to crashes. > Limit what the compiler emits to the oldest machine model currently > supported by Linux. > > Signed-off-by: Nina Schoetterl-Glausch <nsg@linux.ibm.com> > --- Technically it's a fix but as the tests have been running fine for years I'm inclined to put it into a full release rather than a fix rc. Thanks for taking care of the fix: Reviewed-by: Janosch Frank <frankja@linux.ibm.com> > Should we also set -mtune? > Since it are vector instructions that caused the problem here, there > are some alternatives: > * use -mno-vx > * set the required guest control bit to enable vector instructions on > models supporting them > > -march=z10 might prevent similar issues with other instructions, but I > don't know if there actually exist other relevant instructions, so it > could be needlessly restricting. > > > tools/testing/selftests/kvm/Makefile | 3 +++ > 1 file changed, 3 insertions(+) > > diff --git a/tools/testing/selftests/kvm/Makefile b/tools/testing/selftests/kvm/Makefile > index 1750f91dd936..df0989949eb5 100644 > --- a/tools/testing/selftests/kvm/Makefile > +++ b/tools/testing/selftests/kvm/Makefile > @@ -200,6 +200,9 @@ CFLAGS += -Wall -Wstrict-prototypes -Wuninitialized -O2 -g -std=gnu99 \ > -I$(LINUX_TOOL_ARCH_INCLUDE) -I$(LINUX_HDR_PATH) -Iinclude \ > -I$(<D) -Iinclude/$(ARCH_DIR) -I ../rseq -I.. $(EXTRA_CFLAGS) \ > $(KHDR_INCLUDES) > +ifeq ($(ARCH),s390) > + CFLAGS += -march=z10 > +endif > > no-pie-option := $(call try-run, echo 'int main(void) { return 0; }' | \ > $(CC) -Werror $(CFLAGS) -no-pie -x c - -o "$$TMP", -no-pie)
diff --git a/tools/testing/selftests/kvm/Makefile b/tools/testing/selftests/kvm/Makefile index 1750f91dd936..df0989949eb5 100644 --- a/tools/testing/selftests/kvm/Makefile +++ b/tools/testing/selftests/kvm/Makefile @@ -200,6 +200,9 @@ CFLAGS += -Wall -Wstrict-prototypes -Wuninitialized -O2 -g -std=gnu99 \ -I$(LINUX_TOOL_ARCH_INCLUDE) -I$(LINUX_HDR_PATH) -Iinclude \ -I$(<D) -Iinclude/$(ARCH_DIR) -I ../rseq -I.. $(EXTRA_CFLAGS) \ $(KHDR_INCLUDES) +ifeq ($(ARCH),s390) + CFLAGS += -march=z10 +endif no-pie-option := $(call try-run, echo 'int main(void) { return 0; }' | \ $(CC) -Werror $(CFLAGS) -no-pie -x c - -o "$$TMP", -no-pie)
The guest used in s390 kvm selftests is not be set up to handle all instructions the compiler might emit, i.e. vector instructions, leading to crashes. Limit what the compiler emits to the oldest machine model currently supported by Linux. Signed-off-by: Nina Schoetterl-Glausch <nsg@linux.ibm.com> --- Should we also set -mtune? Since it are vector instructions that caused the problem here, there are some alternatives: * use -mno-vx * set the required guest control bit to enable vector instructions on models supporting them -march=z10 might prevent similar issues with other instructions, but I don't know if there actually exist other relevant instructions, so it could be needlessly restricting. tools/testing/selftests/kvm/Makefile | 3 +++ 1 file changed, 3 insertions(+)