On Tue 2021-11-09 17:28:47, Guilherme G. Piccoli wrote: > Currently the "panic_print" parameter/sysctl allows some interesting debug > information to be printed during a panic event. This is useful for example > in cases the user cannot kdump due to resource limits, or if the user > collects panic logs in a serial output (or pstore) and prefers a fast > reboot instead of a kdump. Yes, I have missed this possibility many times. > Happens that currently there's no way to see all CPUs backtraces in > a panic using "panic_print" on architectures that support that. We do > have "oops_all_cpu_backtrace" sysctl, but although partially overlapping > in the functionality, they are orthogonal in nature: "panic_print" is > a panic tuning (and we have panics without oopses, like direct calls to > panic() or maybe other paths that don't go through oops_enter() > function), and the original purpose of "oops_all_cpu_backtrace" is to > provide more information on oopses for cases in which the users desire > to continue running the kernel even after an oops, i.e., used in > non-panic scenarios. panic() already prevents double backtrace of the CPU that Oopsed, see: #ifdef CONFIG_DEBUG_BUGVERBOSE /* * Avoid nested stack-dumping if a panic occurs during oops processing */ if (!test_taint(TAINT_DIE) && oops_in_progress <= 1) dump_stack(); #endif It should be possible to do something similar also for backtraces on all CPUs. There are more situation when the backtraces are printed and panic() is called, for example: softlockup_panic and softlockup_all_cpu_backtrace. Well, it is just nice to have. People probably will not use these options together. And it is better to have the backtraces twice than do not have them at all. > So, we hereby introduce an additional bit for "panic_print" to allow > dumping the CPUs backtraces during a panic event. > > Signed-off-by: Guilherme G. Piccoli <gpiccoli@xxxxxxxxxx> Feel free to use: Reviewed-by: Petr Mladek <pmladek@xxxxxxxx> Best Regards, Petr
