[RFC]: kexec: change to handle memory/cpu changes

[Baoquan He]

On 12/14/20 at 10:50am, Eric DeVolder wrote:
...
> The cell contents show the number of seconds it took for the system to
> process all of the 3840 memblocks. The value in parenthesis is the
> number of kdump unload-then-reload operations per second.
> 
>           1 480GB DIMM   480 1GB DIMMs
> -------+-----------------+----------------+
>  RHEL7 | 181s (21.2 ops) | 389s (9.8 ops) |
> -------+-----------------+----------------+
>  RHEL8 |  86s (44.7 ops) | 419s (9.2 ops) |
> -------+-----------------+----------------+
> 
> The scenario of adding 480 1GiB virtual DIMMs takes more time given
> the larger number of round trips of QEMU -> kernel -> udev -> kernel ->
> QEMU, and are both roughly 400s.
> 
> The RHEL7 system process all 3840 memblocks individually and perform
> 3840 kdump unload-then-reload operations.
> 
> However, RHEL8 data in the best case scenario (1 480GiB DIMM) suggests
> that approximately 86/4= 21 kdump unload-then-reload operations
> happened, and in the worst case scenario (480 1GiB DIMMs), the data
> suggests that approximately 419/4 = 105 kdump unload-then-reload
> operations happened. For RHEL8, the final number of kdump
> unload-then-reload operations are 0.5% (21 of 3840) and 2.7% (105 of
> 3840), respectively, compared to that of the RHEL7 system.
> 
> The throttle approach is quite effective in reducing the number of
> kdump unload-then-reload operations. However, the kdump capture kernel
> is still reloaded multiple times, and each kdump capture kernel reload
> is a race window in which kdump can fail.
> 
> A quick peek at Ubuntu 20.04 LTS reveals it has 50-kdump-tools.rules
> that looks like:
> 
>   SUBSYSTEM=="memory", ACTION=="online", PROGRAM="/usr/sbin/kdump-config try-reload"
>   SUBSYSTEM=="memory", ACTION=="offline", PROGRAM="/usr/sbin/kdump-config try-reload"
>   SUBSYSTEM=="cpu", ACTION=="add", PROGRAM="/usr/sbin/kdump-config try-reload"
>   SUBSYSTEM=="cpu", ACTION=="remove", PROGRAM="/usr/sbin/kdump-config try-reload"
>   SUBSYSTEM=="cpu", ACTION=="offline", PROGRAM="/usr/sbin/kdump-config try-reload"
> 
> which produces the equivalent behavior to RHEL7 whereby every event
> results in a kdump capture kernel reload.
> 
> Fedora 33 and CentOS 8-stream behave the same as RHEL8.
> 
> Perhaps a better solution is to rewrite the vmcoreinfo structure that
> contains the memory and CPU layout information, as those changes to
> memory and CPUs occur. Rewriting vmcoreinfo is an in-kernel activity
> and would certainly avoid the relatively large unload-then-reload
> times of the kdump capture kernel. The pointer to the vmcoreinfo
> structure is provided to the capture kernel via the elfcorehdr=
> parameter to the capture kernel cmdline. Rewriting the vmcoreinfo
> structure as well as rewriting the capture kernel cmdline parameter is
> needed to utilize this approach.

Great investigation and conclusion, and very nice idea as below. When I
read the first half of this mail, I thought maybe we could add a new
option to kexec-tools utility for updating eflcorehdr only when hotplug
udev events detected. Then come to this part, I would say yes, doing it
inside kernel looks better. A special handling for hotplug looks
necessary as you have said, I will check what we can do and give back
some details, thanks for doing these.

Thanks
Baoquan

> 
> Based upon some amount of examining code, I think the challenges
> involved in updating the CPU and memory layout in-kernel are:
> 
>  - adding call-outs on the add_memory()/try_remove_memory() and
>    cpu_up()/cpu_down() paths for notifying the kdump subsystem of
>    memory and/or CPU changes.
> 
>  - updating the struct kimage with the memory or CPU changes
> 
>  - Rewriting the vmcoreinfo structure from the data contained
>    in struct kimage, eg crash_prepare_elf64_headers()
> 
>  - Installing the updated vmcoreinfo struct via
>    kimage_crash_copy_vmcoreinfo() and rewriting the kdump kernel
>    cmdline in order to update parameter elfcorehdr= with the
>    new address
> 
> As I am not overly familiar with all the code paths involved, yet, I'm
> sure the devil is in the details. However, due the kexec_file_load
> syscall, it appears most of the infrastructure is already in place,
> and we essentially need to tap into it again for memory and cpu
> changes.
> 
> It appears that this change could be applicable to both kexec_load and
> kexec_file_load, it has the potential to (eventually) simplify the
> userland kexec utility for kexec_load, and would eliminate the need
> for 98-kexec.rules and the associated churn.
> 
> Comments please!
> eric
> 
> _______________________________________________
> kexec mailing list
> kexec at lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/kexec
>