diff options
Diffstat (limited to 'results/classifier/deepseek-2/output/debug/1647683')
| -rw-r--r-- | results/classifier/deepseek-2/output/debug/1647683 | 35 |
1 files changed, 35 insertions, 0 deletions
diff --git a/results/classifier/deepseek-2/output/debug/1647683 b/results/classifier/deepseek-2/output/debug/1647683 new file mode 100644 index 000000000..818951409 --- /dev/null +++ b/results/classifier/deepseek-2/output/debug/1647683 @@ -0,0 +1,35 @@ + +Bad interaction between tb flushing & gdb stub + +I have been working on a series of patches for ARM big-endian system mode support, using QEMU as a bare-metal simulator for the GDB test suite. At some point I realised that these tests were not running reliably on the QEMU master branch, even without my patches applied. (I.e., in little-endian mode.) + +Running QEMU under GDB in the test harness via Valgrind, using something akin to: + +(gdb) target remote | valgrind --tool=memcheck qemu-arm-system [...] + +leads to intermittent (and quite hard-to-reproduce) segfaults in QEMU of the form: + +==52333== Process terminating with default action of signal 11 (SIGSEGV) +==52333== Access not within mapped region at address 0x24 +==52333== at 0x1D55F2: tb_page_remove (translate-all.c:1026) +==52333== by 0x1D58B4: tb_phys_invalidate (translate-all.c:1119) +==52333== by 0x1D63AA: tb_invalidate_phys_page_range (translate-all.c:1519) +==52333== by 0x1D66D7: tb_invalidate_phys_addr (translate-all.c:1714) +==52333== by 0x1CBA7F: breakpoint_invalidate (exec.c:704) +==52333== by 0x1CC01F: cpu_breakpoint_remove_by_ref (exec.c:869) +==52333== by 0x1CBF97: cpu_breakpoint_remove (exec.c:857) +==52333== by 0x218FAA: gdb_breakpoint_remove (gdbstub.c:717) +==52333== by 0x219E35: gdb_handle_packet (gdbstub.c:1035) +==52333== by 0x21AF62: gdb_read_byte (gdbstub.c:1459) +==52333== by 0x21B096: gdb_chr_receive (gdbstub.c:1672) +==52333== by 0x3AF2BC: qemu_chr_be_write_impl (qemu-char.c:419) + +These crashes didn't happen on a 2.6-era QEMU, so I bisected and discovered the commit 3359baad36889b83df40b637ed993a4b816c4906 ("tcg: Make tb_flush() thread safe") appears to be the thing that triggers this intermittent failure. Reverting the patch on the branch tip makes the crashes go away. + +Unfortunately I don't currently have a way to trigger the segfaults outside of Mentor Graphics's test infrastructure, which I can't share. + +Does anyone know a reason that this might be happening, or suggestions of how I might further debug this? Maybe a missed tb flush in the gdb stub code, somewhere? + +Thanks! + +Julian \ No newline at end of file |