Add an option to create XZ-compressed cores

[jtl]

Summary

Add an option to use multi-threaded XZ compression on kernel cores. On a
machine with a large amount of memory and a large number of CPUs or CPU
cores, this can significantly reduce the amount of time it takes to
create a kernel core.

Obtained from: Netflix

Test Plan

This has already been in use at Netflix for a while. It seems to perform reliably there.

[delphij]

Looks good to me (in general).

Have you considered piping the data to a separate xz(1) process (probably also other compressors) instead of incorporating a wrapper around the library though? Or is this intentional for e.g. performance reasons?

sbin/savecore/savecore_xz.c

802

Personally I'd prefer calloc(1, sizeof(struct xz_cookie)).

[bcr]

OK from manpages. Don't forget to bump the document date when you're committing it.

[jtl]

After the latest core-compression patches (which actually can compress the core dump prior to writing it to swap), is there still a need for this patch? I seem to recall that someone said "yes", but I'd like to have confirmation before adding yet another way to save coredumps.

[markj]

In D10390#310378, @jtl wrote:

After the latest core-compression patches (which actually can compress the core dump prior to writing it to swap), is there still a need for this patch? I seem to recall that someone said "yes", but I'd like to have confirmation before adding yet another way to save coredumps.

I suspect not - the kernel can write out zstd-compressed vmcores now. That's much faster than gzip and also speeds up recovery since savecore has to do less I/O to read the vmcore out of the dump device.