The docker client has historically used Transport.TLSClientConfig to set
the scheme for the API client. A recent moved the resolution to use the
http.Transport directly, rather than save the TLSClientConfig state on a
client struct. This caused issues when mutliple calls made with a single
client would have this field set in the http package on pre-1.7
installations. This fix detects the presence of the TLSClientConfig once
and sets the scheme accordingly.
We still don't know why this issue doesn't happen with Go 1.7 but it
must be more deterministic in the newer version.
Signed-off-by: Stephen J Day <stephen.day@docker.com>
Signed-off-by: Sean Rodman <srodman7689@gmail.com>
Updated the check for the permission error to use os.IsPermission instead of checking the error string. Also, changed the PermissionDenied method to just a new error.
Fixed a typo in client/request.go
Fixed Error name as specified by Pull request builder output.
Worked on making changes to the permissiondenied error.
Fixed typo
Signed-off-by: Sean Rodman <srodman7689@gmail.com>
Updated error message as requested.
Fixed the error as requested
Signed-off-by: Sean Rodman <srodman7689@gmail.com>
Under the convoluted code path for the transport configuration,
TLSConfig was being set even though the socket type is unix. This caused
other code detecting the TLSConfig to assume https, rather than using
the http scheme. This led to a situation where if `DOCKER_CERT_PATH` is
set, unix sockets start reverting to https. There is other odd behavior
from go-connections that is also reproduced here.
For the most part, we try to reproduce the side-effecting behavior from
go-connections to retain the current docker behavior. This whole mess
needs to ripped out and fixed, as this pile spaghetti is unnacceptable.
This code is way to convoluted for an http client. We'll need to fix
this but the Go API will break to do it.
Signed-off-by: Stephen J Day <stephen.day@docker.com>
This package doesn't really seem to do anything of real interest.
Removing it and replacing with a few helper functions. Most of this was
maintaining a fork of ctxhttp to support a mock that was unnecessary.
We could probably do with a further refactor of the client interface.
There is a lot of confusion of between transport, http layer and
application layer that makes for some awkward code. This change
improves the situation to the point where no breaking changes are
introduced.
Signed-off-by: Stephen J Day <stephen.day@docker.com>
Instead of reformatting error from the request action, we wrap it,
allowing the cause to be recovered. This is important for consumers that
need to be able to detect context errors, such as `Cancelled` and
`DeadlineExceeded`.
Signed-off-by: Stephen J Day <stephen.day@docker.com>
John Howard
Stephen J Day
Aaron Lehmann
Sean Rodman
Michael Crosby