Skip to contentSkip to conversation

hw/riscv/sifive_clint.c: avoid integer overflow in timecmp write

PATCH In review Fabien Chouteau 3 participants 5 messages Jan 30, 2019 – Feb 13, 2019

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 hw/riscv/sifive_clint.c: avoid integer overflow in timecmp write

Writing a high value in timecmp leads to an integer overflow. This patch modifies the code to detect such case, and use the maximum integer value as the next trigger for the timer.

---
 hw/riscv/sifive_clint.c | 31 ++++++++++++++++++++++++++-----
 1 file changed, 26 insertions(+), 5 deletions(-)
hw/riscv/sifive_clint.c +26 -5
index d4c159e937..1ca1f8c75e 100644
@@ -38,8 +38,10 @@ static uint64_t cpu_riscv_read_rtc(void)
  */
 static void sifive_clint_write_timecmp(RISCVCPU *cpu, uint64_t value)
 {
-    uint64_t next;
     uint64_t diff;
+    uint64_t lapse_ns;
+    uint64_t clock_ns;
+    int64_t  next_ns;
 
     uint64_t rtc_r = cpu_riscv_read_rtc();
 
@@ -54,10 +56,29 @@ static void sifive_clint_write_timecmp(RISCVCPU *cpu, uint64_t value)
     /* otherwise, set up the future timer interrupt */
     riscv_cpu_update_mip(cpu, MIP_MTIP, BOOL_TO_MASK(0));
     diff = cpu->env.timecmp - rtc_r;
-    /* back to ns (note args switched in muldiv64) */
-    next = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) +
-        muldiv64(diff, NANOSECONDS_PER_SECOND, SIFIVE_CLINT_TIMEBASE_FREQ);
-    timer_mod(cpu->env.timer, next);
+
+    /*
+     * How many nanoseconds until the next trigger (note args switched in
+     * muldiv64)
+     */
+    lapse_ns = muldiv64(diff,
+                        NANOSECONDS_PER_SECOND,
+                        SIFIVE_CLINT_TIMEBASE_FREQ);
+
+    /* Current time in nanoseconds */
+    clock_ns = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL);
+
+    if ((G_MAXINT64 - clock_ns) <= lapse_ns) {
+        /*
+         * clock + lapse would overflow on 64bit. The highest 64bit value is
+         * used as the next trigger time.
+         */
+        next_ns = G_MAXINT64;
+    } else {
+        next_ns = clock_ns + lapse_ns;
+    }
+
+    timer_mod(cpu->env.timer, next_ns);
 }
 
 /*
Signature · 1 line
2.17.1
#1 lore raw
Quoted 45 lines · Fabien Chouteau

Writing a high value in timecmp leads to an integer overflow. This patch modifies the code to detect such case, and use the maximum integer value as the next trigger for the timer.

---
 hw/riscv/sifive_clint.c | 31 ++++++++++++++++++++++++++-----
 1 file changed, 26 insertions(+), 5 deletions(-)
hw/riscv/sifive_clint.c +14 -5
index d4c159e937..1ca1f8c75e 100644
@@ -38,8 +38,10 @@ static uint64_t cpu_riscv_read_rtc(void)
  */
 static void sifive_clint_write_timecmp(RISCVCPU *cpu, uint64_t value)
 {
-    uint64_t next;
     uint64_t diff;
+    uint64_t lapse_ns;
+    uint64_t clock_ns;
+    int64_t  next_ns;
 
     uint64_t rtc_r = cpu_riscv_read_rtc();
 
@@ -54,10 +56,29 @@ static void sifive_clint_write_timecmp(RISCVCPU *cpu, uint64_t value)
     /* otherwise, set up the future timer interrupt */
     riscv_cpu_update_mip(cpu, MIP_MTIP, BOOL_TO_MASK(0));
     diff = cpu->env.timecmp - rtc_r;
-    /* back to ns (note args switched in muldiv64) */
-    next = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) +
-        muldiv64(diff, NANOSECONDS_PER_SECOND, SIFIVE_CLINT_TIMEBASE_FREQ);
-    timer_mod(cpu->env.timer, next);
+
+    /*
+     * How many nanoseconds until the next trigger (note args switched in
+     * muldiv64)
+     */
+    lapse_ns = muldiv64(diff,
+                        NANOSECONDS_PER_SECOND,
+                        SIFIVE_CLINT_TIMEBASE_FREQ);
+
+    /* Current time in nanoseconds */
+    clock_ns = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL);

hw/riscv/sifive_clint.c static void sifive_clint_write_timecmp(RISCVCPU *cpu, uint64_t value) · line 79

Quoted 10 lines
hw/riscv/sifive_clint.c +10 -0
+
+    if ((G_MAXINT64 - clock_ns) <= lapse_ns) {
+        /*
+         * clock + lapse would overflow on 64bit. The highest 64bit value is
+         * used as the next trigger time.
+         */
+        next_ns = G_MAXINT64;
+    } else {
+        next_ns = clock_ns + lapse_ns;
+    }

Can you describe what this fixes?

Won't an overflow be ok as we then just wrap around anyway? I guess there is a problem if we want a value so large that we wrap around past our current time though.

Alistair

Quoted 9 lines
hw/riscv/sifive_clint.c +2 -1
+
+    timer_mod(cpu->env.timer, next_ns);
 }
 
 /*
--

2.17.1

#2 lore raw

Hello Alistair,

On 07/02/2019 01:42, Alistair Francis wrote:>

Quoted 2 lines

Can you describe what this fixes?

I encountered this problem when I tried to write 0xffffffffffffffff in timecmp.

With the integer overflow in QEMU, writing this value means that the QEMU timer will be set in the past.

Quoted 4 lines

Won't an overflow be ok as we then just wrap around anyway? I guess there is a problem if we want a value so large that we wrap around past our current time though.

The overflow was in the computation of the value next_ns. It is used to set the QEMU timer:

timer_mod(cpu->env.timer, next_ns);

A negative next_ns -because of the overflow- means that the timer triggers immediately instead of far in the future.

Regards,

Reviewed-by
#3 lore raw
Quoted 17 lines · Fabien Chouteau

Hello Alistair,

On 07/02/2019 01:42, Alistair Francis wrote:>

Quoted 2 lines

Can you describe what this fixes?

I encountered this problem when I tried to write 0xffffffffffffffff in timecmp.

With the integer overflow in QEMU, writing this value means that the QEMU timer will be set in the past.

Quoted 4 lines

Won't an overflow be ok as we then just wrap around anyway? I guess there is a problem if we want a value so large that we wrap around past our current time though.

Quoted 7 lines

The overflow was in the computation of the value next_ns. It is used to set the QEMU timer:

timer_mod(cpu->env.timer, next_ns);

A negative next_ns -because of the overflow- means that the timer triggers immediately instead of far in the future.

Ah you are right here. The expired time of the timer will be set to zero (it looks like QEMU ensures it can't be negative) but then it detects that as being in the past and will trigger the timer event as timer_expired_ns() will return true.

Alistair

Quoted 2 lines

Regards,

#4 lore raw
Quoted 26 lines · On Fri, 08 Feb 2019 10:41:17 PST (-0800), [email protected] wrote:
Quoted 17 lines · Fabien Chouteau

Hello Alistair,

On 07/02/2019 01:42, Alistair Francis wrote:>

Quoted 2 lines

Can you describe what this fixes?

I encountered this problem when I tried to write 0xffffffffffffffff in timecmp.

With the integer overflow in QEMU, writing this value means that the QEMU timer will be set in the past.

Quoted 4 lines

Won't an overflow be ok as we then just wrap around anyway? I guess there is a problem if we want a value so large that we wrap around past our current time though.

Quoted 7 lines

The overflow was in the computation of the value next_ns. It is used to set the QEMU timer:

timer_mod(cpu->env.timer, next_ns);

A negative next_ns -because of the overflow- means that the timer triggers immediately instead of far in the future.

Quoted 6 lines

Ah you are right here. The expired time of the timer will be set to zero (it looks like QEMU ensures it can't be negative) but then it detects that as being in the past and will trigger the timer event as timer_expired_ns() will return true.

Thanks. I'll target this for the next PR.