Hector Martin

what are the guarantees about what happens if a computer crashes during a write()? is it possible for the file to end up with corrupted data at the end, or is each byte of the write atomic?

I've seen files end up padded with zeroes after an unsynced write and hard shutdown, many times. So either ext4 data=ordered does not work as advertised, isn't enabled by default on some systems, or something else is often broken further down the stack.

big oof

Seriously, "history file corrupt" is a common zsh complaint for me after a hard shutdown (on several systems) and then I have to open it in vim and nuke the last line full of ^@.

I've seen it with text syslogs too. And of course when it happens to systemd-jourmald binary logs it just craps itself hard and refuses to load them.

Ah, here we go, some discussion about this with @tytso. data=ordered only guarantees you won't get *garbage* (e.g. leaked free disk blocks, which is a huge security issue), but you can still get zeroes. https://plus.google.com/+KentonVarda/posts/JDwHfAiLGNQ …

i told you about how macOS's hardware video decode driver used to allocate vmem for frames without zeroing it, right?

I'd say "nice", but really, this is the kind of crap I expect in, like, every proprietary GPU driver for everything. Performance!!!!!ʷʰᵃᵗ ᶦˢ ᵗʰᵃᵗ ˢᵉᶜᵘʳᶦᵗʸ ᵗʰᶦⁿᵍ ʸᵒᵘ ˢᵃʸ?