This is not as necessary as it was for nix, since internal/pkg only unblocks exclusive artifacts one at a time. Still, this is useful when running alongside an unprivileged music player which cannot set itself to a higher priority.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This got brought up earlier today as a potential optimisation. This change documents why it is not viable, and hopefully clears up some performance implications of using IRDecoder, namely that its decoding costs do not amortise.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This significantly improves performance and is a good assumption since the primary use case of FileArtifact is over the network.
Signed-off-by: Ophestra <cat@gensokyo.uk>
TContext no longer validates FileArtifact ahead of time, validation outcome is instead determined after consuming the reader to EOF. All data must therefore be treated as untrusted input until the reader is closed.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This should hopefully provide good separation between the artifact curing backend implementation and the (still work in progress) language. Making the IR parseable also guarantees uniqueness of the representation.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This also rearranges the ident ir to be more predictable, and avoids an obvious and somewhat easy to get into inconsistent state.
Signed-off-by: Ophestra <cat@gensokyo.uk>
No good reason to filter this in the execArtifact container, and the extended filter breaks certain programs.
Signed-off-by: Ophestra <cat@gensokyo.uk>
Any fine-grained file-based locking here significantly hurts performance and is not part of the use case of the package. This change guarantees exclusive access to prevent inconsistent state on the filesystem.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This significantly improves scrubbing performance. Since the cache directory structure is friendly to simultaneous access, this is possible without synchronisation.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This eliminates duplicate identifier computations. The new implementation also significantly reduces allocations while computing identifier for a large dependency tree.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This is significantly more practical than keeping track of them in directory flattening format and setting this in every non-artifact implementation. Only tarArtifact can have meaningful deterministic checksums that are not zero and zeroing them still keeps autotools happy.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This proxies program output through msg with a name and fd prefix. This also avoids introducing additional information to the container via process stdout/stderr.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This does not have a reasonable way of inferring the underlying name. For zero value it falls back to base of executable pathname.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This result was mistakenly unused resulting in incorrect identifiers for artifacts with duplicate dependencies.
Signed-off-by: Ophestra <cat@gensokyo.uk>
An upcoming improvement in the container init makes the current host net check return the same result for both cases. This change
Signed-off-by: Ophestra <cat@gensokyo.uk>
This imposes a hard upper limit to concurrency during dependency satisfaction and moves all dependency-related code out of individual implementations of Artifact. This change also includes ctx and msg as part of Cache.
Signed-off-by: Ophestra <cat@gensokyo.uk>
The cure is condemned at the point of cancellation and all of its state is destroyed by the deferred cleanup, so it makes little sense to wait for it much.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This removes all but the /work/ auto overlay behaviour and enables much greater flexibility. This also renames ExecContainerPath to ExecPath so it is easier to type.
Signed-off-by: Ophestra <cat@gensokyo.uk>
