Am 27.03.2025 um 17:24 hat ger...@altlinux.org geschrieben: > From: Denis Rastyogin <ger...@altlinux.org> > > This error was discovered by fuzzing qemu-img. > > Previously, new I/O requests were launched synchronously inside the > completion callback `bench_cb`, leading to deep recursion and stack > overflow. This patch moves the launching of new requests to a separate > function `bench_bh`, scheduled via `qemu_bh_schedule` to run in the event > loop context, thus unwinding the stack and preventing overflow. > > Signed-off-by: Vasiliy Kovalev <kova...@altlinux.org> > Signed-off-by: Denis Rastyogin <ger...@altlinux.org>
Normally, the callback shouldn't immediately be called, except for errors, which take a different code path anyway. Was this tested with the null block driver or with a backend that is actually relevant in practice? I wonder if switching qemu-img bench to coroutines would make sense. But since this is a benchmarking tool, we need to measure the performance difference from both an additional BH and from switching to coroutines compared to the current state. In case of doubt, I'd leave this unfixed, this isn't something that is run in production. Kevin
