| Message ID | 20200707104205.25085-1-thuth@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [kvm-unit-tests,v2] s390x/cpumodel: The missing DFP facility on TCG is expected | expand |
On 07/07/2020 12.42, Thomas Huth wrote: > When running the kvm-unit-tests with TCG on s390x, the cpumodel test > always reports the error about the missing DFP (decimal floating point) > facility. This is kind of expected, since DFP is not required for > running Linux and thus nobody is really interested in implementing > this facility in TCG. Thus let's mark this as an expected error instead, > so that we can run the kvm-unit-tests also with TCG without getting > test failures that we do not care about. > > Signed-off-by: Thomas Huth <thuth@redhat.com> > --- > v2: > - Rewrote the logic, introduced expected_tcg_fail flag > - Use manufacturer string instead of VM name to detect TCG > > s390x/cpumodel.c | 49 ++++++++++++++++++++++++++++++++++++++++++------ > 1 file changed, 43 insertions(+), 6 deletions(-) > > diff --git a/s390x/cpumodel.c b/s390x/cpumodel.c > index 5d232c6..190ceb2 100644 > --- a/s390x/cpumodel.c > +++ b/s390x/cpumodel.c > @@ -11,14 +11,19 @@ > */ > > #include <asm/facility.h> > +#include <alloc_page.h> > > -static int dep[][2] = { > +static struct { > + int facility; > + int implied; > + bool expected_tcg_fail; > +} dep[] = { > /* from SA22-7832-11 4-98 facility indications */ > { 4, 3 }, > { 5, 3 }, > { 5, 4 }, > { 19, 18 }, > - { 37, 42 }, > + { 37, 42, true }, /* TCG does not have DFP and won't get it soon */ > { 43, 42 }, > { 73, 49 }, > { 134, 129 }, > @@ -38,6 +43,36 @@ static int dep[][2] = { > { 155, 77 }, > }; > > +/* > + * A hack to detect TCG (instead of KVM): QEMU uses "TCGguest" as guest > + * name by default when we are running with TCG (otherwise it's "KVMguest") I forgot to update the comment, it rather should read: "A hack to detect TCG (instead of KVM): Check the manufacturer string which differs between TCG and KVM." (or something similar) Thomas
On Tue, 7 Jul 2020 12:42:05 +0200 Thomas Huth <thuth@redhat.com> wrote: > When running the kvm-unit-tests with TCG on s390x, the cpumodel test > always reports the error about the missing DFP (decimal floating point) > facility. This is kind of expected, since DFP is not required for > running Linux and thus nobody is really interested in implementing > this facility in TCG. Thus let's mark this as an expected error instead, > so that we can run the kvm-unit-tests also with TCG without getting > test failures that we do not care about. > > Signed-off-by: Thomas Huth <thuth@redhat.com> > --- > v2: > - Rewrote the logic, introduced expected_tcg_fail flag > - Use manufacturer string instead of VM name to detect TCG > > s390x/cpumodel.c | 49 ++++++++++++++++++++++++++++++++++++++++++------ > 1 file changed, 43 insertions(+), 6 deletions(-) (...) > +static bool is_tcg(void) > +{ > + const char qemu_ebcdic[] = { 0xd8, 0xc5, 0xd4, 0xe4 }; > + bool ret = false; > + uint8_t *buf; > + > + buf = alloc_page(); > + if (!buf) > + return false; > + > + if (stsi(buf, 1, 1, 1)) { > + goto out; > + } This does an alloc_page() and a stsi() every time you call it... > + > + /* > + * If the manufacturer string is "QEMU" in EBCDIC, then we are on TCG > + * (otherwise the string is "IBM" in EBCDIC) > + */ > + if (!memcmp(&buf[32], qemu_ebcdic, sizeof(qemu_ebcdic))) > + ret = true; > +out: > + free_page(buf); > + return ret; > +} > + > + > int main(void) > { > int i; > @@ -46,11 +81,13 @@ int main(void) > > report_prefix_push("dependency"); ...so maybe cache the value for is_tcg() here instead of checking multiple times in the loop? (Or cache the value in is_tcg().) > for (i = 0; i < ARRAY_SIZE(dep); i++) { > - if (test_facility(dep[i][0])) { > - report(test_facility(dep[i][1]), "%d implies %d", > - dep[i][0], dep[i][1]); > + if (test_facility(dep[i].facility)) { > + report_xfail(dep[i].expected_tcg_fail && is_tcg(), > + test_facility(dep[i].implied), > + "%d implies %d", > + dep[i].facility, dep[i].implied); > } else { > - report_skip("facility %d not present", dep[i][0]); > + report_skip("facility %d not present", dep[i].facility); > } > } > report_prefix_pop();
On 07.07.20 13:44, Cornelia Huck wrote: > On Tue, 7 Jul 2020 12:42:05 +0200 > Thomas Huth <thuth@redhat.com> wrote: > >> When running the kvm-unit-tests with TCG on s390x, the cpumodel test >> always reports the error about the missing DFP (decimal floating point) >> facility. This is kind of expected, since DFP is not required for >> running Linux and thus nobody is really interested in implementing >> this facility in TCG. Thus let's mark this as an expected error instead, >> so that we can run the kvm-unit-tests also with TCG without getting >> test failures that we do not care about. >> >> Signed-off-by: Thomas Huth <thuth@redhat.com> >> --- >> v2: >> - Rewrote the logic, introduced expected_tcg_fail flag >> - Use manufacturer string instead of VM name to detect TCG >> >> s390x/cpumodel.c | 49 ++++++++++++++++++++++++++++++++++++++++++------ >> 1 file changed, 43 insertions(+), 6 deletions(-) > > (...) > >> +static bool is_tcg(void) >> +{ >> + const char qemu_ebcdic[] = { 0xd8, 0xc5, 0xd4, 0xe4 }; >> + bool ret = false; >> + uint8_t *buf; >> + >> + buf = alloc_page(); >> + if (!buf) >> + return false; >> + >> + if (stsi(buf, 1, 1, 1)) { >> + goto out; >> + } > > This does an alloc_page() and a stsi() every time you call it... > >> + >> + /* >> + * If the manufacturer string is "QEMU" in EBCDIC, then we are on TCG >> + * (otherwise the string is "IBM" in EBCDIC) >> + */ >> + if (!memcmp(&buf[32], qemu_ebcdic, sizeof(qemu_ebcdic))) >> + ret = true; >> +out: >> + free_page(buf); >> + return ret; >> +} >> + >> + >> int main(void) >> { >> int i; >> @@ -46,11 +81,13 @@ int main(void) >> >> report_prefix_push("dependency"); > > ...so maybe cache the value for is_tcg() here instead of checking > multiple times in the loop? Maybe move it to common code and do the detection early during boot? The n provide is_tcg() or sth. like that. Could be helpful in other context maybe.
On 7/7/20 1:45 PM, David Hildenbrand wrote: > On 07.07.20 13:44, Cornelia Huck wrote: >> On Tue, 7 Jul 2020 12:42:05 +0200 >> Thomas Huth <thuth@redhat.com> wrote: >> >>> When running the kvm-unit-tests with TCG on s390x, the cpumodel test >>> always reports the error about the missing DFP (decimal floating point) >>> facility. This is kind of expected, since DFP is not required for >>> running Linux and thus nobody is really interested in implementing >>> this facility in TCG. Thus let's mark this as an expected error instead, >>> so that we can run the kvm-unit-tests also with TCG without getting >>> test failures that we do not care about. >>> >>> Signed-off-by: Thomas Huth <thuth@redhat.com> >>> --- >>> v2: >>> - Rewrote the logic, introduced expected_tcg_fail flag >>> - Use manufacturer string instead of VM name to detect TCG >>> >>> s390x/cpumodel.c | 49 ++++++++++++++++++++++++++++++++++++++++++------ >>> 1 file changed, 43 insertions(+), 6 deletions(-) >> >> (...) >> >>> +static bool is_tcg(void) >>> +{ >>> + const char qemu_ebcdic[] = { 0xd8, 0xc5, 0xd4, 0xe4 }; >>> + bool ret = false; >>> + uint8_t *buf; >>> + >>> + buf = alloc_page(); >>> + if (!buf) >>> + return false; >>> + >>> + if (stsi(buf, 1, 1, 1)) { >>> + goto out; >>> + } >> >> This does an alloc_page() and a stsi() every time you call it... >> >>> + >>> + /* >>> + * If the manufacturer string is "QEMU" in EBCDIC, then we are on TCG >>> + * (otherwise the string is "IBM" in EBCDIC) >>> + */ >>> + if (!memcmp(&buf[32], qemu_ebcdic, sizeof(qemu_ebcdic))) >>> + ret = true; >>> +out: >>> + free_page(buf); >>> + return ret; >>> +} >>> + >>> + >>> int main(void) >>> { >>> int i; >>> @@ -46,11 +81,13 @@ int main(void) >>> >>> report_prefix_push("dependency"); >> >> ...so maybe cache the value for is_tcg() here instead of checking >> multiple times in the loop? > > Maybe move it to common code and do the detection early during boot? The > n provide is_tcg() or sth. like that. Could be helpful in other context > maybe. > Well we also already have a check for zvm 6 with stsi 3.2.2 in skey.c I'm not completely convinced that I want to loose two pages and a few cycles on every startup for two separate test cases.
On 07/07/2020 17.09, Janosch Frank wrote: > On 7/7/20 1:45 PM, David Hildenbrand wrote: >> On 07.07.20 13:44, Cornelia Huck wrote: >>> On Tue, 7 Jul 2020 12:42:05 +0200 >>> Thomas Huth <thuth@redhat.com> wrote: >>> >>>> When running the kvm-unit-tests with TCG on s390x, the cpumodel test >>>> always reports the error about the missing DFP (decimal floating point) >>>> facility. This is kind of expected, since DFP is not required for >>>> running Linux and thus nobody is really interested in implementing >>>> this facility in TCG. Thus let's mark this as an expected error instead, >>>> so that we can run the kvm-unit-tests also with TCG without getting >>>> test failures that we do not care about. >>>> >>>> Signed-off-by: Thomas Huth <thuth@redhat.com> >>>> --- >>>> v2: >>>> - Rewrote the logic, introduced expected_tcg_fail flag >>>> - Use manufacturer string instead of VM name to detect TCG >>>> >>>> s390x/cpumodel.c | 49 ++++++++++++++++++++++++++++++++++++++++++------ >>>> 1 file changed, 43 insertions(+), 6 deletions(-) >>> >>> (...) >>> >>>> +static bool is_tcg(void) >>>> +{ >>>> + const char qemu_ebcdic[] = { 0xd8, 0xc5, 0xd4, 0xe4 }; >>>> + bool ret = false; >>>> + uint8_t *buf; >>>> + >>>> + buf = alloc_page(); >>>> + if (!buf) >>>> + return false; >>>> + >>>> + if (stsi(buf, 1, 1, 1)) { >>>> + goto out; >>>> + } >>> >>> This does an alloc_page() and a stsi() every time you call it... >>> >>>> + >>>> + /* >>>> + * If the manufacturer string is "QEMU" in EBCDIC, then we are on TCG >>>> + * (otherwise the string is "IBM" in EBCDIC) >>>> + */ >>>> + if (!memcmp(&buf[32], qemu_ebcdic, sizeof(qemu_ebcdic))) >>>> + ret = true; >>>> +out: >>>> + free_page(buf); >>>> + return ret; >>>> +} >>>> + >>>> + >>>> int main(void) >>>> { >>>> int i; >>>> @@ -46,11 +81,13 @@ int main(void) >>>> >>>> report_prefix_push("dependency"); >>> >>> ...so maybe cache the value for is_tcg() here instead of checking >>> multiple times in the loop? >> >> Maybe move it to common code and do the detection early during boot? The >> n provide is_tcg() or sth. like that. Could be helpful in other context >> maybe. >> > > Well we also already have a check for zvm 6 with stsi 3.2.2 in skey.c > I'm not completely convinced that I want to loose two pages and a few > cycles on every startup for two separate test cases. It certainly does not make sense to run the stsi calls during each and every startup ... but I can move the is_tcg() function to the library and make it a little bit smarter. I'll prepare a v3... Thomas
diff --git a/s390x/cpumodel.c b/s390x/cpumodel.c index 5d232c6..190ceb2 100644 --- a/s390x/cpumodel.c +++ b/s390x/cpumodel.c @@ -11,14 +11,19 @@ */ #include <asm/facility.h> +#include <alloc_page.h> -static int dep[][2] = { +static struct { + int facility; + int implied; + bool expected_tcg_fail; +} dep[] = { /* from SA22-7832-11 4-98 facility indications */ { 4, 3 }, { 5, 3 }, { 5, 4 }, { 19, 18 }, - { 37, 42 }, + { 37, 42, true }, /* TCG does not have DFP and won't get it soon */ { 43, 42 }, { 73, 49 }, { 134, 129 }, @@ -38,6 +43,36 @@ static int dep[][2] = { { 155, 77 }, }; +/* + * A hack to detect TCG (instead of KVM): QEMU uses "TCGguest" as guest + * name by default when we are running with TCG (otherwise it's "KVMguest") + */ +static bool is_tcg(void) +{ + const char qemu_ebcdic[] = { 0xd8, 0xc5, 0xd4, 0xe4 }; + bool ret = false; + uint8_t *buf; + + buf = alloc_page(); + if (!buf) + return false; + + if (stsi(buf, 1, 1, 1)) { + goto out; + } + + /* + * If the manufacturer string is "QEMU" in EBCDIC, then we are on TCG + * (otherwise the string is "IBM" in EBCDIC) + */ + if (!memcmp(&buf[32], qemu_ebcdic, sizeof(qemu_ebcdic))) + ret = true; +out: + free_page(buf); + return ret; +} + + int main(void) { int i; @@ -46,11 +81,13 @@ int main(void) report_prefix_push("dependency"); for (i = 0; i < ARRAY_SIZE(dep); i++) { - if (test_facility(dep[i][0])) { - report(test_facility(dep[i][1]), "%d implies %d", - dep[i][0], dep[i][1]); + if (test_facility(dep[i].facility)) { + report_xfail(dep[i].expected_tcg_fail && is_tcg(), + test_facility(dep[i].implied), + "%d implies %d", + dep[i].facility, dep[i].implied); } else { - report_skip("facility %d not present", dep[i][0]); + report_skip("facility %d not present", dep[i].facility); } } report_prefix_pop();
When running the kvm-unit-tests with TCG on s390x, the cpumodel test always reports the error about the missing DFP (decimal floating point) facility. This is kind of expected, since DFP is not required for running Linux and thus nobody is really interested in implementing this facility in TCG. Thus let's mark this as an expected error instead, so that we can run the kvm-unit-tests also with TCG without getting test failures that we do not care about. Signed-off-by: Thomas Huth <thuth@redhat.com> --- v2: - Rewrote the logic, introduced expected_tcg_fail flag - Use manufacturer string instead of VM name to detect TCG s390x/cpumodel.c | 49 ++++++++++++++++++++++++++++++++++++++++++------ 1 file changed, 43 insertions(+), 6 deletions(-)