This allows for more flexibility during implementation. The use case that required this was for expanding single directory tarballs.
Signed-off-by: Ophestra <cat@gensokyo.uk>
The previous implementation exposes arbitrary user input to the cache as an identifier, which is highly error-prone and can cause the cache to enter an inconsistent state if the user is not careful. This change replaces the implementation to compute identifier late, using url string as params.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This moves all cache I/O code to Cache. Artifact now only contains methods for constructing their actual contents.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This must be writable to enable renaming, and the final result is conventionally read-only alongside the entire directory contents. This change overrides the permission bits as part of Store.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This is still not ideal as it makes entry into Store sequential. This will be improved after more usage code is written.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This works on any directories and should be robust against any bad state the artifact curing process might have failed at.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This provides infrastructure for computing a deterministic identifier based on current artifact kind, opaque parameters data, and optional dependency kind and identifiers.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This makes the decoder safe against untrusted input without hurting performance for a trusted stream. This should still not be called against untrusted input though.
Signed-off-by: Ophestra <cat@gensokyo.uk>
The fact that Gob serialisation is deterministic is an implementation detail. This change replaces Gob with a simple custom format.
Signed-off-by: Ophestra <cat@gensokyo.uk>
Turns out this did not work because in the vm test harness, virtualisation.fileSystems completely and silently overrides fileSystems, causing its contents to not even be evaluated anymore. This is not documented as far as I can tell, and is not obvious by any stretch of the imagination. The current hack is cargo culted from nix-community/impermanence and hopefully lasts until this project fully replaces nix.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This optional behaviour is required on NixOS as it is otherwise impossible to set this up: systemd.mounts breaks startup order somehow even though my unit looks identical to generated ones, fileSystems does not support any kind of initialisation or ordering other than against other mount points.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This should have been handled in a custom option parsing function, but that much extra complexity is unnecessary for this edge case. Honestly I do not know why libfuse does not handle this itself.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This replaces the forking daemonise libfuse function which prevents Go callbacks from calling into the runtime. This also enforces least privilege on the daemon process.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This is not really just library wrapper functions, but instead implements the callbacks, so fuse-operations makes more sense.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This is never called directly anywhere and it is simple enough to be included in the macro. This avoids passing the pointer around and dereferencing errno location, resulting in over 5% increase in throughput on the clang build. No change in the gcc build though.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This filesystem does not support symbolic links, so readlink is not useful, and unreachable in this case because of the check in getattr.
Signed-off-by: Ophestra <cat@gensokyo.uk>
The interface provided by net is not used here and is a leftover from a previous implementation. This change removes it.
Signed-off-by: Ophestra <cat@gensokyo.uk>
The interface does not expose underlying kernel notification mechanisms. This change removes the need to poll in situations were the next call might block.
This is made cumbersome by the SyscallConn interface left over from a previous implementation, it will be replaced in a later commit as the current implementation does not make use of any net.Conn methods other than Close.
Signed-off-by: Ophestra <cat@gensokyo.uk>
When returned wrapped as a syscall error, these are impossible to recover from, so wrap them as a fatal error.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This no longer exposes the pipewire socket to the container, and instead mediates access via pipewire-pulse. This makes insecure parts of the protocol inaccessible as explained in the doc comment in hst.
Closes#29.
Signed-off-by: Ophestra <cat@gensokyo.uk>
This is a private work directory owned by the specific shim. Useful for sockets owned by this instance of the shim and requires no direct assistance from the priv-side process.
Signed-off-by: Ophestra <cat@gensokyo.uk>
