blob: e19a3341f7867014e3a8f5a8f48fb4a15dfce761 (
plain) (
blame)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
|
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
|