I've posted bug report with extended tests (incl. case without sleep=off). You may find guest image (kernel) in bug description. https://bugs.launchpad.net/qemu/+bug/1795369
The most annoying thing is that some issues are almost not reproducible. There are definitely race conditions somewhere in qemu code. Running 'stress-ng' utility with CPU and I/O stressors in parallel with qemu execution greatly minimizes amount of attempts when I'm trying to trigger some of issues I encounter. I'll try 'info monitor' command tomorrow, but no guarantees that I'll be able to reproduce issue again. Speaking about '-nographic' and SDL... I've noted that UI greatly minimizes possibility of hanging (but not avoids it completely) when using icount in general, so this effect isn't rr-specific. I've already reported this bug too. пн, 1 окт. 2018 г., 20:14 dovgaluk <dovga...@ispras.ru>: > Artem Pisarenko писал 2018-09-30 14:01: > > Feature still broken :( > > Thanks for testing. > > > > > Brief description of my tests. > > > > Guest image is Linux, which just powers off after kernel boots > > (instead of proceeding to user-space /init or /sbin/init). > > Base cmdline: > > qemu-system-x86_64 -nodefaults -machine pc,accel=tcg -m 2048 -cpu > > qemu64 -rtc clock=vm,base=2000-01-01T00:00:00 -kernel bzImage -initrd > > rootfs -append 'nokaslr console=ttyS0 rdinit=/init_poweroff' > > -nographic -serial SERIAL_VALUE -icount > > 1,sleep=off,rr=RR_VALUE,rrfile=icount_rr_capture.bin > > I've never tried it with sleep=off. Can you remove it and try again? > > We also seen a problem with '-nographic'. When we remove this option and > QEMU runs with SDL > window, everything is ok. There is some problem with main loop which may > sleep when there > is no GUI to update, or something like that. We couldn't fix it yet. > > > > > Test 1. When SERIAL_VALUE=none > > Running with RR_VALUE=record completes successfully. > > Running with RR_VALUE=replay doesn't completes. qemu process just > > eating ~100% cpu and memory usage doesn't grow after some moment. I > > don't see what happens because of problem no.2 (see below). > > Try 'info replay' monitor command. Does instruction counter increases? > > > > > Test 2. When SERIAL_VALUE=stdio > > Running with RR_VALUE=record completes successfully. > > > > Running with RR_VALUE=replay caues exit with error: > > > > "qemu-system-x86_64: Missing character write event in the replay log" > > > > These problems are same with qemu 2.12 (both vanilla and with previous > > versions of these patches applied). Furthemore, I consider whole > > icount mode broken and determinism isn't achievable. > > The irony is that I actually don't need record/replay feature. I've > > tried to use it only as instrument to debug failing determinism in > > qemu code. But since replay/record feature itself relies on > > determinism, which is broken, it's no wonder why it fails also (I just > > hoped to bypass it). > > > > Contact me if you need more details. I just tired a lot trying to get > > all these things working... Hope is leaving me... > > Can you share the kernel in case the icount still broken? > > > Pavel Dovgalyuk > > -- С уважением, Артем Писаренко
