On Tue 2021-11-09 17:28:48, Guilherme G. Piccoli wrote: > Currently we have the "panic_print" parameter/sysctl to allow some extra > information to be printed in a panic event. On the other hand, the kdump > mechanism allows to kexec a new kernel to collect a memory dump for the > running kernel in case of panic. > Right now these options are incompatible: the user either sets the kdump > or makes use of "panic_print". The code path of "panic_print" isn't > reached when kdump is configured. > > There are situations though in which this would be interesting: for > example, in systems that are very memory constrained, a handcrafted > tiny kernel/initrd for kdump might be used in order to only collect the > dmesg in kdump kernel. Even more common, systems with no disk space for > the full (compressed) memory dump might very well rely in this > functionality too, dumping only the dmesg with the additional information > provided by "panic_print". Is anyone really using this approach? kmsg_dump() looks like a better choice when there are memory constrains. It does not need to reserve memory for booting the crash kernel. I would not mind much but this change depends on a not fully reliable assumption, see below. Also it will also complicate the solution for the kmsg_dump() code path. It would be better to discuss this togeter with the other patch https://lore.kernel.org/r/20220106212835.119409-1-gpiccoli@xxxxxxxxxx > So, this is what the patch does: allows both functionality to co-exist; > if "panic_print" is set and the system performs a kdump, the extra > information is printed on dmesg before the kexec. Some notes about the > design choices here: > > (a) We could have introduced a sysctl or an extra bit on "panic_print" > to allow enabling the co-existence of kdump and "panic_print", but seems > that would be over-engineering; we have 3 cases, let's check how this > patch change things: > > - if the user have kdump set and not "panic_print", nothing changes; > - if the user have "panic_print" set and not kdump, nothing changes; > - if both are enabled, now we print the extra information before kdump, > which is exactly the goal of the patch (and should be the goal of the > user, since they enabled both options). > > (b) We assume that the code path won't return from __crash_kexec() > so we didn't guard against double execution of panic_print_sys_info(). This sounds suspiciously. There is small race window but it actually works. __crash_kexec() really never returns when @kexec_crash_image is loaded. Well, it might break in the future if the code is modified. Best Regards, Petr
