path: root/results/classifier/zero-shot/105/graphic/1859021

graphic: 0.964
device: 0.960
instruction: 0.959
assembly: 0.955
semantic: 0.953
network: 0.950
other: 0.948
socket: 0.947
boot: 0.944
mistranslation: 0.925
vnc: 0.921
KVM: 0.874

qemu-system-aarch64 (tcg):  cval + voff overflow not handled, causes qemu to hang

The Armv8 architecture reference manual states that for any timer set (e.g. CNTP* and CNTV*), the condition for such timer to generate an interrupt (if enabled & unmasked) is:

CVAL <= CNT(P/V)CT

Although this is arguably sloppy coding, I have seen code that is therefore assuming it can set CVAL to a very high value (e.g. UINT64_MAX) and leave the interrupt enabled in CTL, and never get the interrupt.

On latest master commit as the time of writing, there is an integer overflow in target/arm/helper.c gt_recalc_timer affecting the virtual timer when the interrupt is enabled in CTL:

    /* Next transition is when we hit cval */
    nexttick = gt->cval + offset;

When this overflow happens, I notice that qemu is no longer responsive and that I have to SIGKILL the process:
    - qemu takes nearly all the cpu time of the cores it is running on (e.g. 50% cpu usage if running on half the cores) and is completely unresponsive
    - no guest interrupt (reported via -d int) is generated

Here the minimal code example to reproduce the issue:

    mov     x0, #1
    msr     cntvoff_el2, x0
    mov     x0, #-1
    msr     cntv_cval_el0, x0
    mov     x0, #1
    msr     cntv_ctl_el0, x0 // interrupt generation enabled, not masked; qemu will start to hang here

Options used:
-nographic -machine virt,virtualization=on,gic-version=2,accel=tcg -cpu cortex-a57
-smp 4 -m 1024 -kernel whatever.elf -d unimp,guest_errors,int -semihosting-config enable,target=native
-serial mon:stdio

Version used: 4.2

Bug: https://bugs.launchpad.net/bugs/1859021

Signed-off-by: Alex Bennée <email address hidden>
---
 tests/tcg/aarch64/system/vtimer.c         | 48 +++++++++++++++++++++++
 tests/tcg/aarch64/Makefile.softmmu-target |  4 ++
 2 files changed, 52 insertions(+)
 create mode 100644 tests/tcg/aarch64/system/vtimer.c

diff --git a/tests/tcg/aarch64/system/vtimer.c b/tests/tcg/aarch64/system/vtimer.c
new file mode 100644
index 00000000000..42f2f7796c7
--- /dev/null
+++ b/tests/tcg/aarch64/system/vtimer.c
@@ -0,0 +1,48 @@
+/*
+ * Simple Virtual Timer Test
+ *
+ * Copyright (c) 2020 Linaro Ltd
+ *
+ * SPDX-License-Identifier: GPL-2.0-or-later
+ */
+
+#include <inttypes.h>
+#include <minilib.h>
+
+/* grabbed from Linux */
+#define __stringify_1(x...) #x
+#define __stringify(x...)   __stringify_1(x)
+
+#define read_sysreg(r) ({                                           \
+            uint64_t __val;                                         \
+            asm volatile("mrs %0, " __stringify(r) : "=r" (__val)); \
+            __val;                                                  \
+})
+
+#define write_sysreg(r, v) do {                     \
+        uint64_t __val = (uint64_t)(v);             \
+        asm volatile("msr " __stringify(r) ", %x0"  \
+                 : : "rZ" (__val));                 \
+} while (0)
+
+int main(void)
+{
+    int i;
+
+    ml_printf("VTimer Test\n");
+
+    write_sysreg(cntvoff_el2, 1);
+    write_sysreg(cntv_cval_el0, -1);
+    write_sysreg(cntv_ctl_el0, 1);
+
+    ml_printf("cntvoff_el2=%lx\n", read_sysreg(cntvoff_el2));
+    ml_printf("cntv_cval_el0=%lx\n", read_sysreg(cntv_cval_el0));
+    ml_printf("cntv_ctl_el0=%lx\n", read_sysreg(cntv_ctl_el0));
+
+    /* Now read cval a few times */
+    for (i = 0; i < 10; i++) {
+        ml_printf("%d: cntv_cval_el0=%lx\n", i, read_sysreg(cntv_cval_el0));
+    }
+
+    return 0;
+}
diff --git a/tests/tcg/aarch64/Makefile.softmmu-target b/tests/tcg/aarch64/Makefile.softmmu-target
index 7b4eede3f07..62cdddbb215 100644
--- a/tests/tcg/aarch64/Makefile.softmmu-target
+++ b/tests/tcg/aarch64/Makefile.softmmu-target
@@ -62,3 +62,7 @@ run-memory-replay: memory-replay run-memory-record
 	  "$< on $(TARGET_NAME)")
 
 EXTRA_TESTS+=memory-record memory-replay
+
+# vtimer test
+QEMU_EL2_MACHINE=-machine virt,virtualization=on,gic-version=2 -cpu cortex-a57 -smp 4
+run-vtimer: QEMU_OPTS=$(QEMU_EL2_MACHINE) $(QEMU_SEMIHOST)  -kernel
-- 
2.20.1



If we don't detect this we will be stuck in a busy loop as we schedule
a timer for before now which will continually trigger gt_recalc_timer
even though we haven't reached the state required to trigger the IRQ.

Bug: https://bugs.launchpad.net/bugs/1859021
Cc: <email address hidden>
Signed-off-by: Alex Bennée <email address hidden>
---
 target/arm/helper.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/target/arm/helper.c b/target/arm/helper.c
index 19a57a17da5..eb17106f7bd 100644
--- a/target/arm/helper.c
+++ b/target/arm/helper.c
@@ -2481,6 +2481,9 @@ static void gt_recalc_timer(ARMCPU *cpu, int timeridx)
         } else {
             /* Next transition is when we hit cval */
             nexttick = gt->cval + offset;
+            if (nexttick < gt->cval) {
+                nexttick = UINT64_MAX;
+            }
         }
         /* Note that the desired next expiry time might be beyond the
          * signed-64-bit range of a QEMUTimer -- in this case we just
-- 
2.20.1



On Fri, 10 Jan 2020 at 16:16, Alex Bennée <email address hidden> wrote:
>
> If we don't detect this we will be stuck in a busy loop as we schedule
> a timer for before now which will continually trigger gt_recalc_timer
> even though we haven't reached the state required to trigger the IRQ.
>
> Bug: https://bugs.launchpad.net/bugs/1859021
> Cc: <email address hidden>
> Signed-off-by: Alex Bennée <email address hidden>
> ---
>  target/arm/helper.c | 3 +++
>  1 file changed, 3 insertions(+)
>
> diff --git a/target/arm/helper.c b/target/arm/helper.c
> index 19a57a17da5..eb17106f7bd 100644
> --- a/target/arm/helper.c
> +++ b/target/arm/helper.c
> @@ -2481,6 +2481,9 @@ static void gt_recalc_timer(ARMCPU *cpu, int timeridx)
>          } else {
>              /* Next transition is when we hit cval */
>              nexttick = gt->cval + offset;
> +            if (nexttick < gt->cval) {
> +                nexttick = UINT64_MAX;
> +            }
>          }

There's something odd going on with this code. Adding a bit of context:

        uint64_t offset = timeridx == GTIMER_VIRT ?
                                      cpu->env.cp15.cntvoff_el2 : 0;
        uint64_t count = gt_get_countervalue(&cpu->env);
        /* Note that this must be unsigned 64 bit arithmetic: */
        int istatus = count - offset >= gt->cval;
        [...]
        if (istatus) {
            /* Next transition is when count rolls back over to zero */
            nexttick = UINT64_MAX;
        } else {
            /* Next transition is when we hit cval */
            nexttick = gt->cval + offset;
        }

I think this patch is correct, in that the 'nexttick' values
are all absolute and this cval/offset combination implies
that the next timer interrupt is going to be in a future
so distant we can't even fit the duration in a uint64_t.

But the other half of the 'if' also looks wrong: that's
for the case of "timer has fired, how long until the
wraparound causes the interrupt line to go low again?".
UINT64_MAX is right for the EL1 case where offset is 0,
but the offset might actually be set such that the wrap
around happens fairly soon. We want to calculate the
tick when (count - offset) hits 0, saturated to
UINT64_MAX. It's getting late here and I couldn't figure
out what that expression should be with 15 minutes of
fiddling around with pen and paper diagrams. I'll have another
go tomorrow if nobody else gets there first...

thanks
-- PMM


On Thu, 16 Jan 2020 at 18:45, Peter Maydell <email address hidden> wrote:
> There's something odd going on with this code. Adding a bit of context:
>
>         uint64_t offset = timeridx == GTIMER_VIRT ?
>                                       cpu->env.cp15.cntvoff_el2 : 0;
>         uint64_t count = gt_get_countervalue(&cpu->env);
>         /* Note that this must be unsigned 64 bit arithmetic: */
>         int istatus = count - offset >= gt->cval;
>         [...]
>         if (istatus) {
>             /* Next transition is when count rolls back over to zero */
>             nexttick = UINT64_MAX;
>         } else {
>             /* Next transition is when we hit cval */
>             nexttick = gt->cval + offset;
>         }
>
> I think this patch is correct, in that the 'nexttick' values
> are all absolute and this cval/offset combination implies
> that the next timer interrupt is going to be in a future
> so distant we can't even fit the duration in a uint64_t.
>
> But the other half of the 'if' also looks wrong: that's
> for the case of "timer has fired, how long until the
> wraparound causes the interrupt line to go low again?".
> UINT64_MAX is right for the EL1 case where offset is 0,
> but the offset might actually be set such that the wrap
> around happens fairly soon. We want to calculate the
> tick when (count - offset) hits 0, saturated to
> UINT64_MAX. It's getting late here and I couldn't figure
> out what that expression should be with 15 minutes of
> fiddling around with pen and paper diagrams. I'll have another
> go tomorrow if nobody else gets there first...

With a fresher brain:

For the if (istatus) branch we want the absolute tick
when (count - offset) wraps round to 0, saturated to UINT64_MAX.
I think this is:
    if (offset <= count) {
        nexttick = UINT64_MAX;
    } else {
        nexttick = offset;
    }

Should we consider this a separate bugfix to go in its own patch?

thanks
-- PMM


A different approach was posted that basically elides the overflow case by not scheduling timers for IRQ events which have already happened:

  https://lists.gnu.org/archive/html/qemu-devel/2020-07/msg07915.html


This is an automated cleanup. This bug report has been moved
to QEMU's new bug tracker on gitlab.com and thus gets marked
as 'expired' now. Please continue with the discussion here:

 https://gitlab.com/qemu-project/qemu/-/issues/60