Adding Guilherme G. Piccoli into Cc. On Tue 2026-09-08 14:07:16, Zack Rusin wrote: > The core panic path enters a loaded crash kernel before running kmsg > dumpers, so the VMware logger cannot preserve the panic in the host log. > Set crash_kexec_post_notifiers during VMware platform setup. setup_arch() > runs before ordinary core parameters are parsed, so an explicit > crash_kexec_post_notifiers=0 still overrides this default. The 0644 > parameter also remains writable at runtime. > > VMware exposes no capability bit for this behavior, so the default changes > for every VMware guest. Running the logger first adds work to the panic > path and can reduce kdump reliability.
AFAIK, the quality of the notifiers is varying. Running all notifiers might reduce the kdump reliability even more. I do not like much the hack with crash_kexec_post_notifiers. It is an all or nothing option. Also it was introduced as a quick hack so that users could decide what is more important for them. But it is not longer a "user" decision when some platforms enforce the ordering because they depend on the notifier. panic() is problematic and it is about compromises. And we need to balance what is important, what is safe, and what is optional. This is why I suggested to introduce more notifiers some time ago, see https://lore.kernel.org/lkml/YfPxvzSzDLjO5ldp@alley/ Guillermo implemented this, see https://lore.kernel.org/all/[email protected]/ But it has stalled because it touched too many subsystems and it was hard to get an agreement. Maybe, we should start with something simple, and introduce one more panic notifier as a start. It might be called either: + "panic_hypervisor_list" because "crash_kexec_post_notifiers = true" seems to be primary set on hypervisors. But I would rather make it more generic and call it + panic_pre_crash_kexec or panic_pre_kdump because there might be more notifiers which are either 100% safe and useful or are worth the risk before calling crash dump. We could put there x86/vmware notifiers as a start. And we could later move there other important notifiers. How does that sound, please? > A full log transfer uses 1027 low-bandwidth hypercalls; three checkpoint > attempts are bounded at 3077. In vmcall tests, a maximum-size transfer > took at most 7.47 ms; three complete transfers comprising 3081 calls took > at most 21.58 ms. > > The crash report remains suppressed while a crash kernel is loaded because > the host may terminate the VM inside that call. A future report from the > crash kernel or kdump userspace can restore the event after saving the > vmcore. Best Regards, Petr
