Skip to contentSkip to conversation

linux-user/s390x: signal with SIGFPE on compare-and-trap

PATCH Superseded Jonathan Albrecht 2 participants 8 messages Jul 7, 2021 – Jul 9, 2021

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

v2 cover linux-user/s390x: signal with SIGFPE on compare-and-trap

qemu-s390x signals with SIGILL on compare-and-trap instructions. This breaks OpenJDK which expects SIGFPE in its implementation of implicit exceptions.

This patch depends on [PATCH v6 0/2] target/s390x: Fix SIGILL and SIGFPE psw.addr reporting
https://lore.kernel.org/qemu-devel/[email protected]/

Based-on: [email protected]

v1 -> v2:
- Update to latest version of '... psw.addr reporting' patch
- Rebase to master and fix conflicts in tests/tcg/s390x/Makefile.target

Jonathan Albrecht (2):
linux-user/s390x: signal with SIGFPE on compare-and-trap tests/tcg: Test that compare-and-trap raises SIGFPE

 linux-user/s390x/cpu_loop.c     |  19 +++---
 tests/tcg/s390x/Makefile.target |   1 +
 tests/tcg/s390x/trap.c          | 102 ++++++++++++++++++++++++++++++++
 3 files changed, 113 insertions(+), 9 deletions(-)
 create mode 100644 tests/tcg/s390x/trap.c
Signature · 1 line
2.31.1
#1 lore raw

PATCH v2 1/2 linux-user/s390x: signal with SIGFPE on compare-and-trap

Currently when a compare-and-trap instruction is executed, qemu will always raise a SIGILL signal. On real hardware, a SIGFPE is raised.

Change the PGM_DATA case in cpu_loop to follow the behavior in linux kernel /arch/s390/kernel/traps.c.
* Only raise SIGILL if DXC == 0
* If DXC matches an IEEE exception, raise SIGFPE with correct si_code
* Raise SIGFPE with si_code == 0 for everything else

When applied on [email protected], this fixes crashes in the java jdk such as the linked bug.

Resolves: https://gitlab.com/qemu-project/qemu/-/issues/319

---
 linux-user/s390x/cpu_loop.c | 19 ++++++++++---------
 1 file changed, 10 insertions(+), 9 deletions(-)
linux-user/s390x/cpu_loop.c +10 -9
index 22f2e89c62..6e7dfb290a 100644
@@ -106,11 +106,13 @@ void cpu_loop(CPUS390XState *env)
 
             case PGM_DATA:
                 n = (env->fpc >> 8) & 0xff;
-                if (n == 0xff) {
-                    /* compare-and-trap */
+                if (n == 0) {
                     goto do_sigill_opn;
-                } else {
-                    /* An IEEE exception, simulated or otherwise.  */
+                }
+
+                sig = TARGET_SIGFPE;
+                if ((n & 0x03) == 0) {
+                    /* An IEEE exception, simulated or otherwise. */
                     if (n & 0x80) {
                         n = TARGET_FPE_FLTINV;
                     } else if (n & 0x40) {
@@ -121,13 +123,12 @@ void cpu_loop(CPUS390XState *env)
                         n = TARGET_FPE_FLTUND;
                     } else if (n & 0x08) {
                         n = TARGET_FPE_FLTRES;
-                    } else {
-                        /* ??? Quantum exception; BFP, DFP error.  */
-                        goto do_sigill_opn;
                     }
-                    sig = TARGET_SIGFPE;
-                    goto do_signal_pc;
+                } else {
+                    /* compare-and-trap */
+                    n = 0;
                 }
+                goto do_signal_pc;
 
             default:
                 fprintf(stderr, "Unhandled program exception: %#x\n", n);
Signature · 1 line
2.31.1
#2 lore raw
Quoted 10 lines · Jonathan Albrecht
linux-user/s390x/cpu_loop.c +3 -0
+                sig = TARGET_SIGFPE;
+                if ((n & 0x03) == 0) {
+                    /* An IEEE exception, simulated or otherwise. */
                      if (n & 0x80) {
                          n = TARGET_FPE_FLTINV;
                      } else if (n & 0x40) {
@@ -121,13 +123,12 @@ void cpu_loop(CPUS390XState *env)
                          n = TARGET_FPE_FLTUND;
                      } else if (n & 0x08) {
                          n = TARGET_FPE_FLTRES;

linux-user/s390x/cpu_loop.c void cpu_loop(CPUS390XState *env) · line 130

Quoted 10 lines
linux-user/s390x/cpu_loop.c +3 -5
-                    } else {
-                        /* ??? Quantum exception; BFP, DFP error.  */
-                        goto do_sigill_opn;
                      }
-                    sig = TARGET_SIGFPE;
-                    goto do_signal_pc;
+                } else {
+                    /* compare-and-trap */
+                    n = 0;
                  }

Nearly, but not quite. Replace the ??? Quantum exception with n = 0, otherwise si_code will be garbage for that case.

The structure of the kernel code is

  if (n != 0) {
    /* do_fp_trap */
    si_code = 0;
    if ((n & 3) == 0) {
      /* select on bits 6 & 7 */
    }
    raise sigfpe w/ si_code
  } else {
    raise sigill
  }

The comment for compare-and-trap is misleading, because there are lots of entries in "Figure 6-2. Data-exception codes (DXC)" which arrive there and are not compare-and-trap.

As a general comment, I think a single switch over DXC would be cleaner for both kernel and qemu. It seems like giving different si_code for e.g. "0x40 IEEE division by zero" and "0x43 Simulated IEEE division by zero" is actively incorrect.

r~

#3 lore raw
Quoted 13 lines · Richard Henderson
Quoted 12 lines · Jonathan Albrecht
linux-user/s390x/cpu_loop.c +3 -2
+                sig = TARGET_SIGFPE;
+                if ((n & 0x03) == 0) {
+                    /* An IEEE exception, simulated or otherwise. */
                      if (n & 0x80) {
                          n = TARGET_FPE_FLTINV;
                      } else if (n & 0x40) {
@@ -121,13 +123,12 @@ void cpu_loop(CPUS390XState *env)
                          n = TARGET_FPE_FLTUND;
                      } else if (n & 0x08) {
                          n = TARGET_FPE_FLTRES;
-                    } else {
-                        /* ??? Quantum exception; BFP, DFP error.  */
Quoted 9 lines

linux-user/s390x/cpu_loop.c void cpu_loop(CPUS390XState *env) · line 130

Quoted 8 lines
linux-user/s390x/cpu_loop.c +3 -3
-                        goto do_sigill_opn;
                      }
-                    sig = TARGET_SIGFPE;
-                    goto do_signal_pc;
+                } else {
+                    /* compare-and-trap */
+                    n = 0;
                  }

Thanks for the review. I should have a v3 ready shortly.

Quoted 3 lines

Nearly, but not quite. Replace the ??? Quantum exception with n = 0, otherwise si_code will be garbage for that case.

Thx I'll fix that.

Quoted 13 lines

The structure of the kernel code is

  if (n != 0) {
    /* do_fp_trap */
    si_code = 0;
    if ((n & 3) == 0) {
      /* select on bits 6 & 7 */
    }
    raise sigfpe w/ si_code
  } else {
    raise sigill
  }
Quoted 4 lines

The comment for compare-and-trap is misleading, because there are lots of entries in "Figure 6-2. Data-exception codes (DXC)" which arrive there and are not compare-and-trap.

I'll make the comment less specific.

Quoted 5 lines

As a general comment, I think a single switch over DXC would be cleaner for both kernel and qemu. It seems like giving different si_code for e.g. "0x40 IEEE division by zero" and "0x43 Simulated IEEE division by zero" is actively incorrect.

I went over the DXC section and I see what you mean about the si_codes for simulated IEEE exceptions. I'll plan on handling those the same as non-simulated IEEE if no objections. Otherwise all non-IEEE will have si_code == 0 except DXC == 0x00 will still goto do_sigill_opn.

Quoted 2 lines

r~

#4 lore raw
Quoted 8 lines · jonathan.albrecht
Quoted 5 lines

As a general comment, I think a single switch over DXC would be cleaner for both kernel and qemu.  It seems like giving different si_code for e.g. "0x40 IEEE division by zero" and "0x43 Simulated IEEE division by zero" is actively incorrect.

I went over the DXC section and I see what you mean about the si_codes for simulated IEEE exceptions. I'll plan on handling those the same as non-simulated IEEE if no objections.

Only if you plan on submitting a similar patch for the kernel. Otherwise, qemu would not match the kernel abi.

r~

#5 lore raw
Quoted 6 lines · Richard Henderson
Quoted 5 lines · jonathan.albrecht
Quoted 5 lines

As a general comment, I think a single switch over DXC would be cleaner for both kernel and qemu.  It seems like giving different si_code for e.g. "0x40 IEEE division by zero" and "0x43 Simulated IEEE division by zero" is actively incorrect.

Quoted 7 lines
Quoted 3 lines

I went over the DXC section and I see what you mean about the si_codes for simulated IEEE exceptions. I'll plan on handling those the same as non-simulated IEEE if no objections.

Only if you plan on submitting a similar patch for the kernel. Otherwise, qemu would not match the kernel abi.

Thanks for clarifying. In that case, I'll handle simulated IEEE the same as the current kernel.

#6 lore raw

PATCH v2 2/2 tests/tcg: Test that compare-and-trap raises SIGFPE

In reply to Jonathan Albrecht
---
 tests/tcg/s390x/Makefile.target |   1 +
 tests/tcg/s390x/trap.c          | 102 ++++++++++++++++++++++++++++++++
 2 files changed, 103 insertions(+)
 create mode 100644 tests/tcg/s390x/trap.c
tests/tcg/s390x/Makefile.target +1 -0
index 0a5b25c156..d440ecd6f7 100644
@@ -9,6 +9,7 @@ TESTS+=exrl-trtr
 TESTS+=pack
 TESTS+=mvo
 TESTS+=mvc
+TESTS+=trap
 
 # This triggers failures on s390x hosts about 4% of the time
 run-signals: signals
tests/tcg/s390x/trap.c new file +102 -0

102 lines not on this page · raw message

Signature · 1 line
2.31.1