summary refs log tree commit diff stats
path: root/results/classifier/accel-gemma3:12b/kvm/2143
diff options
context:
space:
mode:
Diffstat (limited to 'results/classifier/accel-gemma3:12b/kvm/2143')
-rw-r--r--results/classifier/accel-gemma3:12b/kvm/214339
1 files changed, 39 insertions, 0 deletions
diff --git a/results/classifier/accel-gemma3:12b/kvm/2143 b/results/classifier/accel-gemma3:12b/kvm/2143
new file mode 100644
index 00000000..f33a46a2
--- /dev/null
+++ b/results/classifier/accel-gemma3:12b/kvm/2143
@@ -0,0 +1,39 @@
+
+ladr_match can cause bus error due to unaligned fetch
+Description of problem:
+On a SPARC host system, which does not support unaligned fetches, QEMU sometimes takes a bus error in ladr_match.
+Steps to reproduce:
+1. (see QEMU command line above)
+2. let the system boot
+3.
+Additional information:
+Problem is a hack in ladr_match - hw/net/pcnet.c:635 (present since 2006!):
+
+```
+Core was generated by `./qemu-system-sparc -rtc base=utc,clock=host -vga cg3 -g 1024x768x8 -machine SS'.
+Program terminated with signal SIGKILL, Killed.
+#0  0xffffffff7ec3b178 in ladr_match (size=110, buf=0xffffffff7ffee972 "33", s=0x808f2a20) at ../hw/net/pcnet.c:634
+634         if ((*(hdr->ether_dhost)&0x01) &&
+[Current thread is 632 (LWP    1        )]
+(gdb) list
+629     }
+630     
+631     static inline int ladr_match(PCNetState *s, const uint8_t *buf, int size)
+632     {
+633         struct qemu_ether_header *hdr = (void *)buf;
+634         if ((*(hdr->ether_dhost)&0x01) &&
+635             ((uint64_t *)&s->csr[8])[0] != 0LL) {
+636             uint8_t ladr[8] = {
+637                 s->csr[8] & 0xff, s->csr[8] >> 8,
+638                 s->csr[9] & 0xff, s->csr[9] >> 8,
+(gdb) print &s->csr[8]
+$1 = (uint16_t *) 0x808f4a7c
+```
+The address of s->csr[8], in this case, is on a 4-byte boundary not an 8-byte boundary, so the hack to test for 8 bytes (4 x 16-bit words) being 0 by casting the address up to a pointer to uint64_t and dereferencing it fails.
+
+The data does not seem to be allocated with a deterministic alignment, this failure does not always occur.
+
+A solution to avoid alignment errors could be to test
+```
+  (s->csr[8] | s->csr[9] | s->csr[10] | s->csr[11]) != 0
+```