Without this buffering, the compressor was outputting 64 bytes at a
time to the HTTP stream, which was resulting in absurdly small chunk
sizes and a lot of extra overhead. The buffering restores the chunk size
to 32768 bytes, which matches the behavior with 1.8.2.
Times pushing to a local registry:
1.8.2: 0m18.934s
master: 0m20.564s
master+this commit: 0m17.593s
Fixes: #17038
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
PR #15493 removed compression of layers when pushing them to a V2
registry. This this makes layer uploads larger than they should be.
This commit restores the compression. It uses an io.Pipe to turn the
gzip compressor output Writer into a Reader, so the ReadFrom method can
be used on the BlobWriter (which is very important for avoiding many
PATCH requests per layer).
Fixes#17209Fixes#17038
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
After v1.8.3 layer checksum is used for image ID
validation. Rewriting the checksums on push would
mean that next pulls will get different image IDs
and pulls may fail if its detected that same
manifest digest can now point to new image ID.
Fixes#17178
Signed-off-by: Tonis Tiigi <tonistiigi@gmail.com>
Generate a hash chain involving the image configuration, layer digests,
and parent image hashes. Use the digests to compute IDs for each image
in a manifest, instead of using the remotely specified IDs.
To avoid breaking users' caches, check for images already in the graph
under old IDs, and avoid repulling an image if the version on disk under
the legacy ID ends up with the same digest that was computed from the
manifest for that image.
When a calculated ID already exists in the graph but can't be verified,
continue trying SHA256(digest) until a suitable ID is found.
"save" and "load" are not changed to use a similar scheme. "load" will
preserve the IDs present in the tar file.
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
Ensure that layers are not excluded from manifests based on previous pushes.
Continue skipping pushes on layers which were pushed by a previous tag.
Update push multiple tag tests.
Ensure that each tag pushed exists on the registry and is pullable.
Add output comparison on multiple tag push check.
fixes#15536
Signed-off-by: Derek McGowan <derek@mcgstyle.net> (github: dmcgowan)
The Docker Daemon should send actual actions client ask for to issue tokens,
not all the permissions that client is guaranteed.
Signed-off-by: xiekeyang <xiekeyang@huawei.com>
Previously, its other return value was used even when it returned an
error. This is awkward and goes against the convention. It also could
have resulted in a nil pointer dereference when an error was returned
because of an unknown pool type. This changes the unknown pool type
error to a panic (since the pool types are hardcoded at call sites and
must always be "push" or "pull"), and returns a "found" boolean instead
of an error.
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
The practice of buffering to a tempfile during a pushing contributes massively
to slow V2 push performance perception. The protocol was actually designed to
avoid precalculation, supporting cut-through data push. This means we can
assemble the layer, calculate its digest and push to the remote endpoint, all
at the same time.
This should increase performance massively on systems with slow disks or IO
bottlenecks.
Signed-off-by: Stephen J Day <stephen.day@docker.com>
Some structures use int for sizes and UNIX timestamps. On some
platforms, int is 32 bits, so this can lead to the year 2038 issues and
overflows when dealing with large containers or layers.
Consistently use int64 to store sizes and UNIX timestamps in
api/types/types.go. Update related to code accordingly (i.e.
strconv.FormatInt instead of strconv.Itoa).
Use int64 in progressreader package to avoid integer overflow when
dealing with large quantities. Update related code accordingly.
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
Add a trusted flag to force the cli to resolve a tag into a digest via the notary trust library and pull by digest.
On push the flag the trust flag will indicate the digest and size of a manifest should be signed and push to a notary server.
If a tag is given, the cli will resolve the tag into a digest and pull by digest.
After pulling, if a tag is given the cli makes a request to tag the image.
Use certificate directory for notary requests
Read certificates using same logic used by daemon for registry requests.
Catch JSON syntax errors from Notary client
When an uncaught error occurs in Notary it may show up in Docker as a JSON syntax error, causing a confusing error message to the user.
Provide a generic error when a JSON syntax error occurs.
Catch expiration errors and wrap in additional context.
Signed-off-by: Derek McGowan <derek@mcgstyle.net> (github: dmcgowan)
pushV2Tag already deduplicates layers, but the scope of this
deduplication is only for a particular tag. If we are pushing all tags
in a repository, we may check layers several times. Fix this by moving
the layersSeen map from the pushV2Tag function to the v2Pusher struct.
In addition to avoiding some useless round-trips, this makes the "docker
push" output less confusing. It formerly could contain many repeated
lines like:
124e2127157f: Image already exists
124e2127157f: Image already exists
...
Add test coverage based on the "docker push" output: a hash should not
appear multiple times when pushing multiple tags.
Fixes#14873
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
Aaron Lehmann