---- Message d'origine ----
De : Stefan Thöni stefan.thoeni@gapfruit.com
Thank you for your input. Meanwhile, we have tested several things. First we crafted a runscript that would trigger the error at roughly 50% on these circumstances: (..) We went back to the Genode 23.11 release, on which the same runscript and circumstances would produce the error at roughly 10% of the runs.
Indeed if the 'git bissect' strategy just tweaks the probability of the bug occuring, it sounds like going further down this route might just be a heavy-work-low-reward endeavour.
As I recall, the issue is about fetchurl sometimes failing to download a file, returning an error code, with some suspected interplay with a vfs server, right ? I wonder if one could temporarily tweak the 'fetchurl' source code (or 'vfs' code) so that it drops into debugger/monitor, the very moment the bug occurs ?
eg if( http error code such and such ) drop_to_debugger();
That would work kind of like "core dumps" on Unices of old when they stumbled on an assert(), but much nicer, one could immediately inspect the state of the component, understand how it got in trouble. That's a technique I sometimes used back in BeOS days.
Or...
Revisiting the tracing and Heisenbug issue, it might be possible to find forms of tracing that are _so_ lightweight that they do not fret the "Heisenbug", i.e. they do not make the bug disappear into hiding when tracing is activated. For the "media kit" part of my project I published a sort of 'ring buffer' tracing class that just memcpy()s the log string into a tiny "ring" buffer, so it has really minimum impact. It's a bit tricky to use, one has to "freeze" the program once the bug occurs, so as to be able to inspect the ringbuffer *before* the log lines are clobbered by new logs. But, if the bug is indeed reproducible under such conditions, even once in ten runs (10%), it might allow collecting valuable information.
I'll be following this if you keep us posted anyhow... I sometimes see "packet operation=1 failed" errors from fetchurl when downloading Falkon etc packages, with the latest Genode, but depot_download reswpawns it and fetchurl works on the second try, so it's a low priority item in my case.
PS. my procedural 2 cents: if you file a ticket it will be easier to re-visit the history, make suggestions etc, compared to e-mails... Plus, this appears to be an upstream bug, and I believe the Genode team likes to keep abreast of upstream bugs ^^
Cedric