Skip to contentSkip to conversation

Fixing infinite loop on SCLP READ SCP INFO error

PATCH Superseded Pierre Morel 3 participants 10 messages May 30, 2023 – Jun 1, 2023

Subsystems (2)

v3 of 5 Superseded Latest is v5 →

lore
Share & tools
lore
#0 lore raw

v3 cover Fixing infinite loop on SCLP READ SCP INFO error

Aborting on SCLP READ SCP INFO error leads to a deadloop.

The loop is:
abort() -> exit() -> smp_teardown() -> smp_query_num_cpus() -> sclp_get_cpu_num() -> assert() -> abort()

Since smp_setup() is done after sclp_read_info() inside setup() this loop only happens when only the start processor is running. Let sclp_get_cpu_num() return 1 in this case.

Also implement the SCLP_RC_INSUFFICIENT_SCCB_LENGTH handling and repeat the sclp command.

Pierre Morel (2):
s390x: sclp: consider monoprocessor on read_info error
s390x: sclp: Implement SCLP_RC_INSUFFICIENT_SCCB_LENGTH

 lib/s390x/sclp.c | 64 ++++++++++++++++++++++++++++++++++++++++++------
 1 file changed, 56 insertions(+), 8 deletions(-)
Signature · 10 lines
2.31.1

since v2:

- use tabs in first patch
  (Nico)

- Added comments

- Added SCLP_RC_INSUFFICIENT_SCCB_LENGTH handling
#1 lore raw

PATCH v3 1/2 s390x: sclp: consider monoprocessor on read_info error

A kvm-unit-test would hang if an abort happens before SCLP Read SCP Information has completed if sclp_get_cpu_num() does not report at least one CPU.
Since we obviously have one, report it.

---
 lib/s390x/sclp.c | 6 ++++++
 1 file changed, 6 insertions(+)
lib/s390x/sclp.c +6 -0
index 12919ca..34a31da 100644
@@ -121,6 +121,12 @@ int sclp_get_cpu_num(void)
 {
 	if (read_info)
 		return read_info->entries_cpu;
+	/*
+	 * If we fail here and read_info has not being set,
+	 * it means we failed early and we try to abort the test.
+	 * We need to return at least one CPU, and obviously we have
+	 * at least one, for the smp_teardown to correctly work.
+	 */
 	return 1;
 }
 
Signature · 1 line
2.31.1
Reviewed-by on v3 1/2
#2 lore raw

Re: [kvm-unit-tests PATCH v3 1/2] s390x: sclp: consider monoprocessor on read_info error

Quoting Pierre Morel (2023-05-30 14:40:55)

Quoted 4 lines

A kvm-unit-test would hang if an abort happens before SCLP Read SCP Information has completed if sclp_get_cpu_num() does not report at least one CPU.
Since we obviously have one, report it.

Sorry for complaining again, in a discussion with Janosch we found that the description and commit below can be easily misunderstood. I suggest the following wording in the commit description:

s390x: sclp: treat system as single processor when read_info is NULL

When a test abort()s before SCLP read info is completed, the assertion on read_info in sclp_read_info() will fail. Since abort() eventually calls smp_teardown() which in turn calls sclp_get_cpu_num(), this will cause an infinite abort() chain, causing the test to hang.

Fix this by considering the system single processor when read_info is missing.

[...]

Quoted 5 lines
lib/s390x/sclp.c +0 -0
index 12919ca..34a31da 100644

lib/s390x/sclp.c int sclp_get_cpu_num(void) · line 129

Quoted 9 lines
lib/s390x/sclp.c +6 -0
@@ -121,6 +121,12 @@ int sclp_get_cpu_num(void)
 {
        if (read_info)
                return read_info->entries_cpu;
+       /*
+        * If we fail here and read_info has not being set,
+        * it means we failed early and we try to abort the test.
+        * We need to return at least one CPU, and obviously we have
+        * at least one, for the smp_teardown to correctly work.
+        */

Please make this:

Don't abort here if read_info is NULL since abort() calls smp_teardown() which eventually calls this function and thus causes an infinite abort() chain, causing the test to hang. Since we obviously have at least one CPU, just return one.

With these changes:

Sorry for the back and forth.

#3 lore raw

Re: [kvm-unit-tests PATCH v3 1/2] s390x: sclp: consider monoprocessor on read_info error

Quoted 11 lines · Nico Boehr

Quoting Pierre Morel (2023-05-30 14:40:55)

Quoted 4 lines

A kvm-unit-test would hang if an abort happens before SCLP Read SCP Information has completed if sclp_get_cpu_num() does not report at least one CPU.
Since we obviously have one, report it.

Sorry for complaining again, in a discussion with Janosch we found that the description and commit below can be easily misunderstood. I suggest the following wording in the commit description:

s390x: sclp: treat system as single processor when read_info is NULL

Quoted 8 lines

When a test abort()s before SCLP read info is completed, the assertion on read_info in sclp_read_info() will fail. Since abort() eventually calls smp_teardown() which in turn calls sclp_get_cpu_num(), this will cause an infinite abort() chain, causing the test to hang.

Fix this by considering the system single processor when read_info is missing.

[...]

better, I take it

thx

Quoted 26 lines
Quoted 5 lines
lib/s390x/sclp.c +0 -0
index 12919ca..34a31da 100644

lib/s390x/sclp.c int sclp_get_cpu_num(void) · line 129

Quoted 9 lines
lib/s390x/sclp.c +6 -0
@@ -121,6 +121,12 @@ int sclp_get_cpu_num(void)
  {
         if (read_info)
                 return read_info->entries_cpu;
+       /*
+        * If we fail here and read_info has not being set,
+        * it means we failed early and we try to abort the test.
+        * We need to return at least one CPU, and obviously we have
+        * at least one, for the smp_teardown to correctly work.
+        */

Please make this:

Don't abort here if read_info is NULL since abort() calls smp_teardown() which eventually calls this function and thus causes an infinite abort() chain, causing the test to hang. Since we obviously have at least one CPU, just return one.

With these changes:

Sorry for the back and forth.

#4 lore raw

PATCH v3 2/2 s390x: sclp: Implement SCLP_RC_INSUFFICIENT_SCCB_LENGTH

In reply to Pierre Morel

If SCLP_CMDW_READ_SCP_INFO fails due to a short buffer, retry with a greater buffer.

---
 lib/s390x/sclp.c | 58 +++++++++++++++++++++++++++++++++++++++++-------
 1 file changed, 50 insertions(+), 8 deletions(-)
lib/s390x/sclp.c +50 -8
index 34a31da..9d51ca4 100644
@@ -17,13 +17,14 @@
 #include "sclp.h"
 #include <alloc_phys.h>
 #include <alloc_page.h>
+#include <asm/facility.h>
 
 extern unsigned long stacktop;
 
 static uint64_t storage_increment_size;
 static uint64_t max_ram_size;
 static uint64_t ram_size;
-char _read_info[PAGE_SIZE] __attribute__((__aligned__(PAGE_SIZE)));
+char _read_info[2 * PAGE_SIZE] __attribute__((__aligned__(PAGE_SIZE)));
 static ReadInfo *read_info;
 struct sclp_facilities sclp_facilities;
 
@@ -89,10 +90,41 @@ void sclp_clear_busy(void)
 	spin_unlock(&sclp_lock);
 }
 
-static void sclp_read_scp_info(ReadInfo *ri, int length)
+static bool sclp_read_scp_info_extended(unsigned int command, ReadInfo *ri)
+{
+	int cc;
+
+	if (!test_facility(140)) {
+		report_abort("S390_FEAT_EXTENDED_LENGTH_SCCB missing");
+		return false;
+	}
+	if (ri->h.length > (2 * PAGE_SIZE)) {
+		report_abort("SCLP_READ_INFO expected size too big");
+		return false;
+	}
+
+	sclp_mark_busy();
+	memset(&ri->h, 0, sizeof(ri->h));
+	ri->h.length = 2 * PAGE_SIZE;
+
+	cc = sclp_service_call(command, ri);
+	if (cc) {
+		report_abort("SCLP_READ_INFO error");
+		return false;
+	}
+	if (ri->h.response_code != SCLP_RC_NORMAL_READ_COMPLETION) {
+		report_abort("SCLP_READ_INFO error %02x", ri->h.response_code);
+		return false;
+	}
+
+	return true;
+}
+
+static void sclp_read_scp_info(ReadInfo *ri)
 {
 	unsigned int commands[] = { SCLP_CMDW_READ_SCP_INFO_FORCED,
 				    SCLP_CMDW_READ_SCP_INFO };
+	int length = PAGE_SIZE;
 	int i, cc;
 
 	for (i = 0; i < ARRAY_SIZE(commands); i++) {
@@ -101,19 +133,29 @@ static void sclp_read_scp_info(ReadInfo *ri, int length)
 		ri->h.length = length;
 
 		cc = sclp_service_call(commands[i], ri);
-		if (cc)
-			break;
-		if (ri->h.response_code == SCLP_RC_NORMAL_READ_COMPLETION)
+		if (cc) {
+			report_abort("SCLP_READ_INFO error");
 			return;
-		if (ri->h.response_code != SCLP_RC_INVALID_SCLP_COMMAND)
+		}
+
+		switch (ri->h.response_code) {
+		case SCLP_RC_NORMAL_READ_COMPLETION:
+			return;
+		case SCLP_RC_INVALID_SCLP_COMMAND:
 			break;
+		case SCLP_RC_INSUFFICIENT_SCCB_LENGTH:
+			sclp_read_scp_info_extended(commands[i], ri);
+			return;
+		default:
+			report_abort("READ_SCP_INFO failed");
+			return;
+		}
 	}
-	report_abort("READ_SCP_INFO failed");
 }
 
 void sclp_read_info(void)
 {
-	sclp_read_scp_info((void *)_read_info, SCCB_SIZE);
+	sclp_read_scp_info((void *)_read_info);
 	read_info = (ReadInfo *)_read_info;
 }
 
Signature · 1 line
2.31.1
#5 lore raw

Re: [kvm-unit-tests PATCH v3 2/2] s390x: sclp: Implement SCLP_RC_INSUFFICIENT_SCCB_LENGTH

Quoted 4 lines · Pierre Morel

If SCLP_CMDW_READ_SCP_INFO fails due to a short buffer, retry with a greater buffer.

You've been testing using all possible cpus, haven't you?

lib/s390x/sclp.c void sclp_clear_busy(void) · line 98

Quoted 9 lines
  }
lib/s390x/sclp.c +6 -1
-static void sclp_read_scp_info(ReadInfo *ri, int length)
+static bool sclp_read_scp_info_extended(unsigned int command, ReadInfo *ri)
+{
+	int cc;
+
+	if (!test_facility(140)) {
+		report_abort("S390_FEAT_EXTENDED_LENGTH_SCCB missing");

That's the QEMU name for the facility, isn't it?
"extended-length-SCCB facility is missing" might be better since that's the name that the architecture specifies for that feature.

lib/s390x/sclp.c void sclp_clear_busy(void) · line 101

Quoted 3 lines
lib/s390x/sclp.c +3 -0
+		return false;
+	}
+	if (ri->h.length > (2 * PAGE_SIZE)) {

sizeof() would reduce the locations that we have to touch if we ever want to increase the buffer in the future.

lib/s390x/sclp.c void sclp_clear_busy(void) · line 108

Quoted 7 lines
lib/s390x/sclp.c +7 -0
+		report_abort("SCLP_READ_INFO expected size too big");
+		return false;
+	}
+
+	sclp_mark_busy();
+	memset(&ri->h, 0, sizeof(ri->h));
+	ri->h.length = 2 * PAGE_SIZE;

Same here

Quoted 44 lines
lib/s390x/sclp.c +28 -4
+
+	cc = sclp_service_call(command, ri);
+	if (cc) {
+		report_abort("SCLP_READ_INFO error");
+		return false;
+	}
+	if (ri->h.response_code != SCLP_RC_NORMAL_READ_COMPLETION) {
+		report_abort("SCLP_READ_INFO error %02x", ri->h.response_code);
+		return false;
+	}
+
+	return true;
+}
+
+static void sclp_read_scp_info(ReadInfo *ri)
  {
  	unsigned int commands[] = { SCLP_CMDW_READ_SCP_INFO_FORCED,
  				    SCLP_CMDW_READ_SCP_INFO };
+	int length = PAGE_SIZE;
  	int i, cc;
 
  	for (i = 0; i < ARRAY_SIZE(commands); i++) {
@@ -101,19 +133,29 @@ static void sclp_read_scp_info(ReadInfo *ri, int length)
  		ri->h.length = length;
 
  		cc = sclp_service_call(commands[i], ri);
-		if (cc)
-			break;
-		if (ri->h.response_code == SCLP_RC_NORMAL_READ_COMPLETION)
+		if (cc) {
+			report_abort("SCLP_READ_INFO error");
  			return;
-		if (ri->h.response_code != SCLP_RC_INVALID_SCLP_COMMAND)
+		}
+
+		switch (ri->h.response_code) {
+		case SCLP_RC_NORMAL_READ_COMPLETION:
+			return;
+		case SCLP_RC_INVALID_SCLP_COMMAND:
  			break;
+		case SCLP_RC_INSUFFICIENT_SCCB_LENGTH:
+			sclp_read_scp_info_extended(commands[i], ri);
+			return;
+		default:

lib/s390x/sclp.c static void sclp_read_scp_info(ReadInfo *ri, int length) · line 116

Quoted 10 lines
lib/s390x/sclp.c +3 -2
+			report_abort("READ_SCP_INFO failed");
+			return;
+		}
  	}
-	report_abort("READ_SCP_INFO failed");
  }
 
  void sclp_read_info(void)
  {
-	sclp_read_scp_info((void *)_read_info, SCCB_SIZE);

Why did you remove that?
You could have re-tried with the extended-length in sclp_read_scp_info(). Or you could return the rc and introduce a tiny function that tries both lengths depending on the rc.

Quoted 4 lines
lib/s390x/sclp.c +1 -0
+	sclp_read_scp_info((void *)_read_info);
  	read_info = (ReadInfo *)_read_info;
  }
 
#6 lore raw

Re: [kvm-unit-tests PATCH v3 2/2] s390x: sclp: Implement SCLP_RC_INSUFFICIENT_SCCB_LENGTH

Quoted 7 lines · Janosch Frank
Quoted 4 lines · Pierre Morel

If SCLP_CMDW_READ_SCP_INFO fails due to a short buffer, retry with a greater buffer.

You've been testing using all possible cpus, haven't you?

yes up to 248

Quoted 13 lines

lib/s390x/sclp.c void sclp_clear_busy(void) · line 98

Quoted 10 lines

  }
  -static void sclp_read_scp_info(ReadInfo *ri, int length) +static bool sclp_read_scp_info_extended(unsigned int command, ReadInfo *ri)

lib/s390x/sclp.c +5 -0
+{
+    int cc;
+
+    if (!test_facility(140)) {
+        report_abort("S390_FEAT_EXTENDED_LENGTH_SCCB missing");

That's the QEMU name for the facility, isn't it?
"extended-length-SCCB facility is missing" might be better since that's the name that the architecture specifies for that feature.

yes

Quoted 6 lines
Quoted 4 lines

+        return false;
+    }
+    if (ri->h.length > (2 * PAGE_SIZE)) {

sizeof() would reduce the locations that we have to touch if we ever want to increase the buffer in the future.

yes

Quoted 10 lines
Quoted 8 lines

+        report_abort("SCLP_READ_INFO expected size too big"); +        return false;
+    }
+
+    sclp_mark_busy();
+    memset(&ri->h, 0, sizeof(ri->h));
+    ri->h.length = 2 * PAGE_SIZE;

Same here

OK

Quoted 48 lines
Quoted 48 lines
lib/s390x/sclp.c +15 -0
+
+    cc = sclp_service_call(command, ri);
+    if (cc) {
+        report_abort("SCLP_READ_INFO error");
+        return false;
+    }
+    if (ri->h.response_code != SCLP_RC_NORMAL_READ_COMPLETION) {
+        report_abort("SCLP_READ_INFO error %02x", ri->h.response_code);
+        return false;
+    }
+
+    return true;
+}
+
+static void sclp_read_scp_info(ReadInfo *ri)

  {
      unsigned int commands[] = { SCLP_CMDW_READ_SCP_INFO_FORCED, SCLP_CMDW_READ_SCP_INFO };
+    int length = PAGE_SIZE;
      int i, cc;
        for (i = 0; i < ARRAY_SIZE(commands); i++) {

lib/s390x/sclp.c +15 -4
@@ -101,19 +133,29 @@ static void sclp_read_scp_info(ReadInfo *ri, 
 int length)
           ri->h.length = length;
             cc = sclp_service_call(commands[i], ri);
-        if (cc)
-            break;
-        if (ri->h.response_code == SCLP_RC_NORMAL_READ_COMPLETION)
+        if (cc) {
+            report_abort("SCLP_READ_INFO error");
               return;
-        if (ri->h.response_code != SCLP_RC_INVALID_SCLP_COMMAND)
+        }
+
+        switch (ri->h.response_code) {
+        case SCLP_RC_NORMAL_READ_COMPLETION:
+            return;
+        case SCLP_RC_INVALID_SCLP_COMMAND:
               break;
+        case SCLP_RC_INSUFFICIENT_SCCB_LENGTH:
+            sclp_read_scp_info_extended(commands[i], ri);
+            return;
+        default:
+            report_abort("READ_SCP_INFO failed");
+            return;
+        }
       }
Quoted 8 lines

lib/s390x/sclp.c static void sclp_read_scp_info(ReadInfo *ri, int length) · line 116

Quoted 5 lines
lib/s390x/sclp.c +0 -2
-    report_abort("READ_SCP_INFO failed");
   }
     void sclp_read_info(void)
   {
-    sclp_read_scp_info((void *)_read_info, SCCB_SIZE);

Why did you remove that?
You could have re-tried with the extended-length in sclp_read_scp_info(). Or you could return the rc and introduce a tiny function that tries both lengths depending on the rc.

Yes, I can let it here. I found it has little sense to give the length as parameter.

Retrying with extended length in sclp_read_scp_info() is what is done isn'it?

It does not change a lot to let the first used size here so I will let it here.

Quoted 5 lines
Quoted 4 lines
lib/s390x/sclp.c +1 -0
+    sclp_read_scp_info((void *)_read_info);
       read_info = (ReadInfo *)_read_info;
   }
#7 lore raw

Re: [kvm-unit-tests PATCH v3 2/2] s390x: sclp: Implement SCLP_RC_INSUFFICIENT_SCCB_LENGTH

In reply to Janosch Frank

Quoting Janosch Frank (2023-06-01 10:03:06)

Quoted 5 lines · Pierre Morel
Quoted 4 lines

If SCLP_CMDW_READ_SCP_INFO fails due to a short buffer, retry with a greater buffer.

Janosch, I think it makes sense if Pierre picks up Claudios suggestion from here:
https://lore.kernel.org/all/20230530173544.378a63c6@p-imbrenda/

Do you agree?

#8 lore raw

Re: [kvm-unit-tests PATCH v3 2/2] s390x: sclp: Implement SCLP_RC_INSUFFICIENT_SCCB_LENGTH

Quoted 10 lines · Nico Boehr

Quoting Janosch Frank (2023-06-01 10:03:06)

Quoted 5 lines · Pierre Morel
Quoted 4 lines

If SCLP_CMDW_READ_SCP_INFO fails due to a short buffer, retry with a greater buffer.

Janosch, I think it makes sense if Pierre picks up Claudios suggestion from here:
https://lore.kernel.org/all/20230530173544.378a63c6@p-imbrenda/

Do you agree?

from my side:

It simplifies greatly the code and tested without problem.

The documentation says the SCCB length is "at least"... so we can use a greater size from the beginning.

#9 lore raw

Re: [kvm-unit-tests PATCH v3 2/2] s390x: sclp: Implement SCLP_RC_INSUFFICIENT_SCCB_LENGTH

Quoted 11 lines · Pierre Morel
Quoted 9 lines · Nico Boehr

Quoting Janosch Frank (2023-06-01 10:03:06)

Quoted 5 lines · Pierre Morel
Quoted 4 lines

If SCLP_CMDW_READ_SCP_INFO fails due to a short buffer, retry with a greater buffer.

Janosch, I think it makes sense if Pierre picks up Claudios suggestion from here:
https://lore.kernel.org/all/20230530173544.378a63c6@p-imbrenda/

Quoted 10 lines
Quoted 1 line

Do you agree?

from my side:

It simplifies greatly the code and tested without problem.

The documentation says the SCCB length is "at least"... so we can use a greater size from the beginning.

Sounds good