Skip to contentSkip to conversation

x86/APIC: handle overflow in TMICT calculation

PATCH In review Jan Beulich 3 participants 4 messages Apr 9 – Apr 10

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86/APIC: handle overflow in TMICT calculation

With an expiry value on the order of 20 hours, and with a bus scale value of 256k (as supplied by qemu), the (signed) multiplication will be UB. As we've checked that the value is positive, we mean unsigned multiplication anyway. Yet let's play safe against even larger expiry and bus scale values, leveraging the compiler builtin that there is for this purpose.

While there also drop the stray cast from the actual TMICT write.

Fixes 9062553a0dc1 ("added time and accurate timer support")
Fixes b95beb185810 ("x86: Clean up APIC local timer handling")

---
Sadly neither gcc5 nor gcc15 properly optimize the (effectively) two uses of the 0xffffffffU constant: Both use a 2nd register to load the constant (really 0xfffffffeU unless <= is used) a 2nd time.

xen/arch/x86/apic.c +9 -3
@@ -1224,10 +1224,16 @@ int reprogram_timer(s_time_t timeout)
     }
 
     if ( timeout && ((expire = timeout - NOW()) > 0) )
-        apic_tmict = min_t(uint64_t, (bus_scale * expire) >> BUS_SCALE_SHIFT,
-                           UINT32_MAX);
+    {
+        unsigned long product;
 
-    apic_write(APIC_TMICT, (unsigned long)apic_tmict);
+        apic_tmict = UINT32_MAX;
+        if ( !__builtin_umull_overflow(bus_scale, expire, &product) &&
+             (product >>= BUS_SCALE_SHIFT) < apic_tmict )
+            apic_tmict = product;
+    }
+
+    apic_write(APIC_TMICT, apic_tmict);
 
     return apic_tmict || !timeout;
 }
Acked-by on v1 1/1
#1 lore raw
Quoted 6 lines · Jan Beulich

With an expiry value on the order of 20 hours, and with a bus scale value of 256k (as supplied by qemu), the (signed) multiplication will be UB. As we've checked that the value is positive, we mean unsigned multiplication anyway. Yet let's play safe against even larger expiry and bus scale values, leveraging the compiler builtin that there is for this purpose.

Quoted 6 lines

While there also drop the stray cast from the actual TMICT write.

Fixes 9062553a0dc1 ("added time and accurate timer support")
Fixes b95beb185810 ("x86: Clean up APIC local timer handling")
Quoted 17 lines

---
Sadly neither gcc5 nor gcc15 properly optimize the (effectively) two uses of the 0xffffffffU constant: Both use a 2nd register to load the constant (really 0xfffffffeU unless <= is used) a 2nd time.

xen/arch/x86/apic.c +2 -3
@@ -1224,10 +1224,16 @@ int reprogram_timer(s_time_t timeout)
     }
 
     if ( timeout && ((expire = timeout - NOW()) > 0) )
-        apic_tmict = min_t(uint64_t, (bus_scale * expire) >> BUS_SCALE_SHIFT,
-                           UINT32_MAX);
+    {
+        unsigned long product;
 
-    apic_write(APIC_TMICT, (unsigned long)apic_tmict);

xen/arch/x86/apic.c int reprogram_timer(s_time_t timeout) · line 1239

Quoted 10 lines
xen/arch/x86/apic.c +7 -0
+        apic_tmict = UINT32_MAX;
+        if ( !__builtin_umull_overflow(bus_scale, expire, &product) &&
+             (product >>= BUS_SCALE_SHIFT) < apic_tmict )
+            apic_tmict = product;
+    }
+
+    apic_write(APIC_TMICT, apic_tmict);
 
     return apic_tmict || !timeout;
 }

This is fine for staging, but be aware it cannot be backported before 4.21 due to the toolchain baseline (and nothing in CI will notice, I don't think.)

~Andrew

#2 lore raw
Quoted 7 lines · Andrew Cooper
Quoted 6 lines · Jan Beulich

With an expiry value on the order of 20 hours, and with a bus scale value of 256k (as supplied by qemu), the (signed) multiplication will be UB. As we've checked that the value is positive, we mean unsigned multiplication anyway. Yet let's play safe against even larger expiry and bus scale values, leveraging the compiler builtin that there is for this purpose.

Quoted 8 lines
Quoted 6 lines

While there also drop the stray cast from the actual TMICT write.

Fixes 9062553a0dc1 ("added time and accurate timer support")
Fixes b95beb185810 ("x86: Clean up APIC local timer handling")

Thanks.

Quoted 20 lines
Quoted 20 lines
xen/arch/x86/apic.c +9 -3
@@ -1224,10 +1224,16 @@ int reprogram_timer(s_time_t timeout)
     }
 
     if ( timeout && ((expire = timeout - NOW()) > 0) )
-        apic_tmict = min_t(uint64_t, (bus_scale * expire) >> BUS_SCALE_SHIFT,
-                           UINT32_MAX);
+    {
+        unsigned long product;
 
-    apic_write(APIC_TMICT, (unsigned long)apic_tmict);
+        apic_tmict = UINT32_MAX;
+        if ( !__builtin_umull_overflow(bus_scale, expire, &product) &&
+             (product >>= BUS_SCALE_SHIFT) < apic_tmict )
+            apic_tmict = product;
+    }
+
+    apic_write(APIC_TMICT, apic_tmict);
 
Quoted 6 lines

xen/arch/x86/apic.c int reprogram_timer(s_time_t timeout) · line 1239

Quoted 2 lines
xen/arch/x86/apic.c +0 -0
     return apic_tmict || !timeout;
 }

This is fine for staging, but be aware it cannot be backported before 4.21 due to the toolchain baseline (and nothing in CI will notice, I don't think.)

I'm debating with myself whether to replace by an asm() there. (If we expected further uses of those overflow built-ins, we could consider adding non-built- in fallbacks in those older branches. Yet unless something like this was needed in an XSA, it would be solely 4.20 to gain such.)

Luckily in this case I think I would notice myself, as by default I'm building the older trees with gcc 4.8.5 and 7.4.

Jan

Tested-by on v1 1/1
#3 lore raw
In reply to Jan Beulich
Quoted 1 line · Jan Beulich

With an expiry value on the order of 20 hours, and with a bus scale value

If I did my math correctly, I believe 10 hours is sufficient to trigger signed multiplication overflow. 20 hours would result in unsigned multiplication overflow.

Quoted 11 lines

of 256k (as supplied by qemu), the (signed) multiplication will be UB. As we've checked that the value is positive, we mean unsigned multiplication anyway. Yet let's play safe against even larger expiry and bus scale values, leveraging the compiler builtin that there is for this purpose.

While there also drop the stray cast from the actual TMICT write.

Fixes 9062553a0dc1 ("added time and accurate timer support")
Fixes b95beb185810 ("x86: Clean up APIC local timer handling")

Thanks!