Lennart Poettering

hmm, i think i'd prefer if we'd make this a more generic thing, i.e. add a macro similat to the existing JSON_BUILD_PAIR_IN_ADDR_NON_NULL() that adds the address in both formats.

Before we add all this super complex mess, can we please spend some time on investigating on how verity setup performs if we forego libcryptsetup/libdevmapper? My understanding is that we...

> This is a profiling run where dm-verity devices are all already pre-setup so the dm library doesn't have to do anything at all with uevents or other kernel setups...

hmm, also looking at the graph (which i don't really grok, as mentioned), I wonder, how much influence a simple "vmtouch" before the reload would already do. Because if this...

> Sorry, but I can't really follow how libdevmapper syncronization optimizations can help, given _entirely skipping it_ via verity sharing doesn't help either. Maybe I'm missing something? not sure i...

i mean, the only thing we do is probe the fstype in the no-gpt case. but we could in theory optimize that away i guess, i.e. and allow encoding this...

> Unfortunately experiments with reading ahead images all failed to provide any meaningful improvements how can that be? i mean, if this is slow io, and you cache the relevant...

> It still does a lot, with blkid for example, and then mounting, and then reading os-release and extension-release, and so on and so forth. All together it costs more...

so my guess hence: if you say that warming up the images beforehand didn't work, then could this simply be because we set up the loopback images to use direct...

> This works, but it seemed quite ugly to me, and I thought you'd never want to consider it? Are you sure you would prefer that approach? i am not...