This reverts commit 84b2162c1a.
The intent of this commit was to set an idle timeout on a HTTP
connection. If a read took more than 60 seconds to complete, or a write
took more than 60 seconds to complete, the connection would be
considered dead.
This doesn't work properly, because the HTTP internals apparently read
from the connection concurrently while writing. An upload that doesn't
complete in 60 seconds leads to a timeout.
Fixes#19967
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
`Upload` already closes the reader returned by `compress` and the
progressreader passed into it, before returning. But even so, the
io.Copy inside compress' goroutine needs to attempt a read from the
progressreader to notice that it's closed, and this read has a side
effect of outputting a progress message. If this happens after `Upload`
returns, it can result in a write to a closed channel. Change `compress`
to return a channel that allows the caller to wait for its goroutine to
finish before freeing any resources connected to the reader that was
passed to it.
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
Otherwise, some operations can get stuck indefinitely when the remote
side is unresponsive.
Fixes#12823
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
A watcher would output the current progress item when it was detached,
in case it missed that item earlier, which would leave the user seeing
some intermediate step of the operation. This commit changes it to only
output it on detach if it didn't already output the same item.
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
Currently, the temporary file storing downloaded layer data is only
removed after a successful download or a digest verification error. A
transport-level error does not cause it to be removed. This is a
regression from 1.9 that could cause disk usage to grow until the Docker
daemon is restarted.
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
Things could go wrong if Watch was called after the last existing
watcher was released. The call to Watch would succeed even though it was
not really adding a watcher, and the corresponding call to Release would
close hasWatchers a second time.
The fix for this is twofold:
1. We allow transfers to gain new watchers after the watcher count has
touched zero. This means that the channel returned by Released should
not be closed until all watchers have been released AND the transfer is
no longer tracked by the transfer manager, meaning it won't be possible
for additional calls to Watch to race with closing the channel returned
by Released.
The Transfer interface has a new method called Close so the transfer can
know when the transfer manager no longer references it.
Remove the Cancel method. It's not used and should not be exported.
2. Even though (1) makes it possible to add watchers after all the
previous watchers have been released, we want to avoid doing this in
practice. A transfer that has had all its watchers released is in the
process of being cancelled, and attaching to one of these will never be
the correct behavior. Add a check if a watcher is attaching to a
cancelled transfer. In this case, wait for the transfer to be removed
from the map and try again. This will ensure correct behavior when a
watcher tries to attach during the race window.
Either (1) or (2) should be sufficient to fix the race involved here,
but the combination is the most correct approach. (1) fixes the
low-level plumbing to be resilient to the race condition, and (2) avoids
using it in a racy way.
Fixes#19606
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
There was already a check that prevented protocol-level fallback in this
situation, but retries within a specific protocol will still happen.
This makes it take a long time for the pull to finally error out.
This fixes slowness in TestDaemonNoSpaceleftOnDeviceError, which used to
take a long time due to the backoff between retry attempts:
PASS: docker_cli_daemon_test.go:1868: DockerDaemonSuite.TestDaemonNoSpaceleftOnDeviceError 5.882s
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
When authorization errors are returned by the token process the error will be wrapped in url.Error.
In order to check the underlying error for retry this error message should be unwrapped.
Unwrapping this error allows failure to push due to an unauthorized response to keep from retrying, possibly resulting in later 429 responses.
Signed-off-by: Derek McGowan <derek@mcgstyle.net> (github: dmcgowan)
Revert the portions of #17617 that report all errors when a pull
falls back, and go back to just reporting the last error. This was nice
to have, but causes some UX issues because nonexistent images show
additional "unauthorized" errors.
Keep the part of the PR that handled ENOSPC, as this appears to work
even without tracking multiple errors.
Fixes#19419
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
Tracks source repository information for each blob in the blobsum
service, which is then used to attempt to mount blobs from another
repository when pushing instead of having to re-push blobs to the same
registry.
Signed-off-by: Brian Bland <brian.bland@docker.com>
One of the things this test checks is that the progress indicator
completes for each download. Some progress messages may be lost because
only one message is buffered for each download, but the last progress
message is guaranteed not to be lost. The test therefore checks for a
10/10 progress indication.
However, the assumption that this is the last progress message to be
sent is incorrect. The last message is actually "Pull complete". So
check for this instead.
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
A manifest list refers to platform-specific manifests. This allows
for images that target more than one architecture to share the same tag.
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
Currently this always uses the schema1 manifest builder. Later, it will
be changed to attempt schema2 first, and fall back when necessary.
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
The trust code used to parse the console output of `docker push` to
extract the digest, tag, and size information and determine what to
sign. This is fragile and might give an attacker control over what gets
signed if the attacker can find a way to influence what gets printed as
part of the push output.
This commit sends the push metadata out-of-band. It introduces an `Aux`
field in JSONMessage that can carry application-specific data alongside
progress updates. Instead of parsing formatted output, the client looks
in this field to get the digest, size, and tag from the push.
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
- Stop serializing JSONMessage in favor of events.Message.
- Keep backwards compatibility with JSONMessage for container events.
Signed-off-by: David Calavera <david.calavera@gmail.com>
Support restoreCustomImage for windows with a new interface to extract
the graph driver from the LayerStore.
Signed-off-by: Daniel Nephin <dnephin@docker.com>
This is a followup to #18839. That PR relaxed the fallback logic so that
if a manifest doesn't exist on v2, or the user is unauthorized to access
it, we try again with the v1 protocol. A similar special case is needed
for "pull all tags" (docker pull -a). If the v2 registry doesn't
recognize the repository, or doesn't allow the user to access it, we
should fall back to v1 and try to pull all tags from the v1 registry.
Conversely, if the v2 registry does allow us to list the tags, there
should be no fallback, even if there are errors pulling those tags.
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
RWLayer will now have more operations and be protected through a referenced type rather than always looked up by string in the layer store.
Separates creation of RWLayer (write capture layer) from mounting of the layer.
This allows mount labels to be applied after creation and allowing RWLayer objects to have the same lifespan as a container without performance regressions from requiring mount.
Signed-off-by: Derek McGowan <derek@mcgstyle.net> (github: dmcgowan)
PR #18590 caused compatibility issues with registries such as gcr.io
which support both the v1 and v2 protocols, but do not provide the same
set of images over both protocols. After #18590, pulls from these
registries would never use the v1 protocol, because of the
Docker-Distribution-Api-Version header indicating that v2 was supported.
Fix the problem by making an exception for the case where a manifest is
not found. This should allow fallback to v1 in case that image is
exposed over the v1 protocol but not the v2 protocol.
This avoids the overly aggressive fallback behavior before #18590 which
would allow protocol fallback after almost any error, but restores
interoperability with mixed v1/v2 registry setups.
Fixes#18832
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
- Use layer DiffIDs for progress output in v1 push. This makes the
output consistent with v2 pushes, which means that a fallback to v1
won't start progress bars for a different set of IDs.
- Change wording used in v1 status updates to be consistent with v2.
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
If we detect a Docker-Distribution-Api-Version header indicating that
the registry speaks the V2 protocol, no fallback to V1 should take
place.
The same applies if a V2 registry operation succeeds while attempting a
push or pull.
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
This commit adds a transfer manager which deduplicates and schedules
transfers, and also an upload manager and download manager that build on
top of the transfer manager to provide high-level interfaces for uploads
and downloads. The push and pull code is modified to use these building
blocks.
Some benefits of the changes:
- Simplification of push/pull code
- Pushes can upload layers concurrently
- Failed downloads and uploads are retried after backoff delays
- Cancellation is supported, but individual transfers will only be
cancelled if all pushes or pulls using them are cancelled.
- The distribution code is decoupled from Docker Engine packages and API
conventions (i.e. streamformatter), which will make it easier to split
out.
This commit also includes unit tests for the new distribution/xfer
package. The tests cover 87.8% of the statements in the package.
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
The v1 push code was querying the size of the layer chain up to the
layer it was pushing, rather than just that layer. This made the
progress indicator inaccurate.
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
We were calling Stat for each layer to get the size so we could indicate
progress, but https://github.com/docker/distribution/pull/1226 made it
possible to get the length from the GET request that Open initiates.
Saving one round-trip per layer should make pull operations slightly
faster and more robust.
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
This change allows API clients to retrieve an authentication token from
a registry, and then pass that token directly to the API.
Example usage:
REPO_USER=dhiltgen
read -s PASSWORD
REPO=privateorg/repo
AUTH_URL=https://auth.docker.io/token
TOKEN=$(curl -s -u "${REPO_USER}:${PASSWORD}" "${AUTH_URL}?scope=repository:${REPO}:pull&service=registry.docker.io" |
jq -r ".token")
HEADER=$(echo "{\"registrytoken\":\"${TOKEN}\"}"|base64 -w 0 )
curl -s -D - -H "X-Registry-Auth: ${HEADER}" -X POST "http://localhost:2376/images/create?fromImage=${REPO}"
Signed-off-by: Daniel Hiltgen <daniel.hiltgen@docker.com>
Currently, the resources associated with the io.Reader returned by
TarStream are only freed when it is read until EOF. This means that
partial uploads or exports (for example, in the case of a full disk or
severed connection) can leak a goroutine and open file. This commit
changes TarStream to return an io.ReadCloser. Resources are freed when
Close is called.
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
