s390x/cpumodel: The missing DFP facility on TCG is expected

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 <[email protected]>

---
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(-)

s390x/cpumodel.c +7 -2

index 5d232c6..190ceb2 100644
@@ -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 },

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 <[email protected]>

---
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(-)

s390x/cpumodel.c +13 -0

+
+	/*
+	 * 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;

s390x/cpumodel.c +6 -4

 	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();

Quoted 16 lines · On Tue, 7 Jul 2020 12:42:05 +0200 Thomas Huth <[email protected]> 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 <[email protected]>

---
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(-)

(...)

s390x/cpumodel.c static int dep[][2] = { · line 62

Quoted 13 lines

s390x/cpumodel.c +13 -0

+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...

Quoted 18 lines

s390x/cpumodel.c +13 -0

+
+	/*
+	 * 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)
 

s390x/cpumodel.c int main(void) · line 82

Quoted 1 line

s390x/cpumodel.c +0 -0

 	report_prefix_push("dependency");

...so maybe cache the value for is_tcg() here instead of checking multiple times in the loop?

Quoted 56 lines · Cornelia Huck

Quoted 16 lines · On Tue, 7 Jul 2020 12:42:05 +0200 Thomas Huth <[email protected]> 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 <[email protected]>

---
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(-)

(...)

s390x/cpumodel.c static int dep[][2] = { · line 62

Quoted 13 lines

s390x/cpumodel.c +13 -0

+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...

Quoted 18 lines

s390x/cpumodel.c +13 -0

+
+	/*
+	 * 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)
 

Quoted 4 lines

s390x/cpumodel.c int main(void) · line 82

Quoted 1 line

s390x/cpumodel.c +0 -0

 	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.

Quoted 62 lines · David Hildenbrand

Quoted 60 lines · Cornelia Huck

Quoted 16 lines · On Tue, 7 Jul 2020 12:42:05 +0200 Thomas Huth <[email protected]> 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 <[email protected]>

---
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(-)

(...)

s390x/cpumodel.c static int dep[][2] = { · line 62

Quoted 13 lines

s390x/cpumodel.c +13 -0

+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...

Quoted 17 lines

s390x/cpumodel.c +13 -0

+
+	/*
+	 * 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;

s390x/cpumodel.c int main(void) · line 82

Quoted 2 lines

s390x/cpumodel.c +0 -0

@@ -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?

Quoted 4 lines

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.