Our newly added metadata files in the
tests/repository_data/fishy_rolenames/metadata directory have an expiry
date until "2021-10-22T11:21:56Z" and today while running the tests on
develop branch I recived this error:
ExpiredMetadataError("Metadata X expired on Fri Oct 22 11:21:56 2021")
when running the tests in tests/test_updater.py file and more precisly
the TestUpdaterRolenames.test_unusual_rolenames() test.
That's why I decided to bump the expiration date to a random time in
the future (October 22-nd 2050) and I had to resign all of the metadata
files.
Signed-off-by: Martin Vrachev <mvrachev@vmware.com>
For users of legacy client (tuf/client/) this is purely a security fix
release with no API or functionality changes. For ngclient and Metadata
API, some API changes are included.
All users are advised to upgrade.
Signed-off-by: Jussi Kukkonen <jkukkonen@vmware.com>
_fileinfo_has_changed() and _update_fileinfo() have been unused internal
methods since 2016. Remove them.
Signed-off-by: Jussi Kukkonen <jkukkonen@vmware.com>
The original commit 051b8229 handled the loading and saving metadata
cases but the legacy client actually checks for the files existence
in various other places:
* _update_versioninfo() never reads the file but operates differently
depending on whether the file exists or not
* _move_current_to_previous() that copies files around
* MultiRepoUpdater initialization: this only handle root.json so
is still correct
* _update_fileinfo() which is dead code
Fix the first two of these cases.
Make sure rolenames like "../a" won't trick ngclient into creating the
metadata file outside the metadata cache.
The test data was semi-manually created with RepositorySimulator:
this test code could use RepositorySimulator directly instead (like the
ngclient tests do) but that would require some more infrastructural
work.
Signed-off-by: Jussi Kukkonen <jkukkonen@vmware.com>
If an attacker manages to create arbitrary rolenames they could trick
the client into writing metadata files into unexpected locations:
To avoid directory traversal and writing files into unexpected
locations, encode the rolename before using it as filename.
If a client has delegated targets metadata with rolenames that have
percent-encoded characters in them, these metadata will now not be
found in local metadata cache and must be re-downloaded.
Note that this does not mean using rolenames that get encoded is
advisable (as forming the download URLs still has issues with them),
this just means the client will not do unsafe writes when it encounters
rolenames like this.
Signed-off-by: Jussi Kukkonen <jkukkonen@vmware.com>
Add support for adding delegated targets into RepositorySimulator.
Make the metadata URL parsing in RepositorySimulator more robust.
Add a test to make sure "../a" won't trick ngclient into creating the
metadata file outside the metadata cache.
Signed-off-by: Jussi Kukkonen <jkukkonen@vmware.com>
If an attacker manages to create arbitrary rolenames they could trick
the client into writing metadata files into unexpected locations:
To avoid directory traversal and writing files into unexpected
locations, encode the rolename before using it as filename.
If a client has delegated targets metadata with rolenames that have
percent-encoded characters in them, these metadata will now not be
found in local metadata cache and must be re-downloaded.
Note that this does not mean using rolenames that get encoded is
advisable (as forming the download URLs still has issues with them),
this just means the client will not do unsafe writes when it encounters
rolenames like this.
Signed-off-by: Jussi Kukkonen <jkukkonen@vmware.com>
urljoin considers the second URL to override the base URL if the second
one contains e.g. hostname: this could lead to ngclient downloading
from the wrong host entirely. Doing that would not compromise the
security of the system as the metadata would still need to be verified,
but would definitely be unexpected and a bug.
Note that we're still not encoding the rolename, it's just inserted into
the URL as is.
Signed-off-by: Jussi Kukkonen <jkukkonen@vmware.com>
Modify RepositorySimulator function delegates() to all_targets(), so
that all targets can be traversed and updated with one cycle when
calling update_snapshot() (which is the only use case for now for
delegates()).
Signed-off-by: Martin Vrachev <mvrachev@vmware.com>
Add an option to calculate the hashes and length for timestamp/snapshot
meta.
This will help to cover more use cases with the repository simulator.
Signed-off-by: Martin Vrachev <mvrachev@vmware.com>
If you do the following steps:
1. call Updater.refresh() and load, verify and cache all metadata files
2. modify timestamp snapshot meta information:
(One or more of hashes or length for snapshot changes here)
3. call Updater.refresh() again
4. root and timestamp will be updated to their latest versions
5. local snapshot will be loaded, but hashes/length will be different
than the ones in timestamp.snapshot_meta and that will prevent loading
6. remote snapshot is loaded and verification starts
then when executing step 6 the rollback checks will not be done because
the old snapshot was not loaded on step 5.
In order to resolve this issue, we are introducing the idea of trusted and
untrusted snapshot.
Trusted snapshot is the locally available cached version. This version has
been verified at least once meaning hashes and length were already checked
against timestamp.snapshot_meta hashes and length.
That's why we can allow loading a trusted snapshot version even if there is a
mismatch between the current timestamp.snapshot_meta hashes/length and
hashes/length inside the trusted snapshot.
Untrusted snapshot is the one downloaded from the web. It hasn't been verified
before and that's why we mandate that timestamp.snapshot_meta hashes and length
should match the hashes and legth calculated on this untrusted version of
snapshot.
As the TrustedMetadataSet doesn't have information which snapshot is trusted or
not, so possibly the best solution is to add a new argument "trusted"
to update_snapshot.
Even though this is ugly as the rest of the update functions doesn't
have such an argument, it seems the best solution as it seems to work
in all cases:
- when loading a local snapshot, we know the data has at some point been
trusted (signatures have been checked): it doesn't need to match hashes
now
- if there is no local snapshot and we're updating from remote, the
remote data must match meta hashes in timestamp
- if there is a local snapshot and we're updating from remote, the remote
data must match meta hashes in timestamp
Lastly, I want to point out that hash checks for metadata files are not
essential to TUF security guarantees: they are just an additional layer of
security that allows us to avoid even parsing json that could be malicious -
we already know the malicious metadata would be stopped at metadata
verification after the parsing.
Signed-off-by: Martin Vrachev <mvrachev@vmware.com>
The purpose of this config was to ensure blocking
inside the download loop and releasing CPU resources.
To our best knowledge the network stack currently used
in RequestsFetcher will always block inside the loop
and the issue cannot be reproduced.
'chunk_size' and 'socket_timeout' are currently the
settings provided by RequestsFetcher to tweak
CPU usage and download granularity.
Signed-off-by: Teodora Sechkova <tsechkova@vmware.com>
Reuse the decorator defined in tests/utils.py in order
to receive more helpful messages when an assertion
fails in test_tragets().
Signed-off-by: Martin Vrachev <mvrachev@vmware.com>
Generalize the decorator used in test_metadata_serialization.py and
move it inside tests/utils.py, so it can be reused in other similar
situations.
Signed-off-by: Martin Vrachev <mvrachev@vmware.com>
The definition of consistent targets in the spec is ambiguous:
"consistent target files should be written to non-volatile storage
as digest.filename.ext"
Additionally, the specification describes consistent targets when the
client builds the download URL as follows:
"The filename is of the form HASH.FILENAME.EXT".
The issue is about how we interpreted those quotes.
The legacy updater has decided this means a target path "a/b" will
translate to a download url path "a/{HASH}.b".
The ngclient however translates the target path "a/b" to a download url
path "{HASH}.a/b".
We decided we want to follow the same approach taken from the legacy
updater and thus change how we construct the consistent targets.
Additionally, we want to make sure we test for cases when the TARGETPATH
is an empty string or points to a directory.
Signed-off-by: Martin Vrachev <mvrachev@vmware.com>
Doing so is not always safe and has various other issues
(like target paths "a/../b" and "b" ending up as the same
local path).
Instead URL-encode the target path to make it a plain filename. This
removes any opportunity for path trickery and removes the need to create
the required sub directories (which we were not doing currently, leading
to failed downloads). URL-encoding encodes much more than we really need
but doing so should not hurt: the important thing is that it encodes
all path separators.
Return the actual filepath as return value. I would like to modify the
arguments so caller could decide the filename if they want to. But I
won't do it now because updated_targets() (the caching mechanism)
relies on filenames being chosen by TUF. The plan is to make it
possible for caller to choose the filename though.
This is clearly a "filesystem API break" for anyone depending on the
actual target file names, and does not make sense if we do not plan to
go forward with other updated_targets()/download_target() changes
listed in #1580.
This is part of bigger plan in #1580Fixes#1571
Signed-off-by: Jussi Kukkonen <jkukkonen@vmware.com>
The handling of consistent snapshot was not very clear: try to make
it more obvious what is supported and what is not.
Signed-off-by: Jussi Kukkonen <jkukkonen@vmware.com>
* Add very simple targets support into simulator
* Add documentation for the simulator
* Add an example targets test
This might need to be tweaked and/or extended as we add tests but the
implementation should give a good indication of how to extend it.
As an example, non-consistent targets are not yet supported, but
making fetch() check for the consistent_snapshot state and respond
accordingly should be easy.
Signed-off-by: Jussi Kukkonen <jkukkonen@vmware.com>
This may have been required by a linter at some point, but isn't
anymore: Not annotating makes the documentation look better.
Signed-off-by: Jussi Kukkonen <jkukkonen@vmware.com>
Also add a summary to the page -- unfortunately getting a standard
TOC would require creating a rst page for each class.
Signed-off-by: Jussi Kukkonen <jkukkonen@vmware.com>
This makes the individual pages easier to read.
Use some autodoc configuration so we can have less config
in the automodule/autoclass declarations.
Signed-off-by: Jussi Kukkonen <jkukkonen@vmware.com>
Situation before
* constructor args are not documented
* object attributes are documented
* sphinx cannot show object attribute type annotations
* attribute docs take a lot of vertical space
Now:
* constructor args are documented
* sphinx can show annotated types of constructor args
* class docstring now explains the attributes are the same as
constructor args (and attributes are not explicitly documented)
Signed-off-by: Jussi Kukkonen <jkukkonen@vmware.com>
Instead of starting a dedicated slow_retrieval_server
to test for read timeout in RequestsFetcher, use
unittest.mock to mock the response.raw.read call.
Signed-off-by: Teodora Sechkova <tsechkova@vmware.com>
RequestsFetcher now handles connect/read timeout
errors on session.get() as it does when reading data.
Adding a test which uses unittest.mock to patch the
session.get method instead of simulating the timeout
with a server.
Signed-off-by: Teodora Sechkova <tsechkova@vmware.com>
This is an API change to the exceptions thrown in Root.add_key()
and Root.remove_key().
The reason for that change is that in my opinion the correct exceptions
in these cases should be "ValueError" instead of "KeyError" as
the problems are in the given values - role doesn't exist or
key is not used by a particular role.
Additionally, document the thrown exceptions in "Root.add_key" and
add a test which invokes that exception.
Signed-off-by: Martin Vrachev <mvrachev@vmware.com>
Root class has the functionality to add and remove keys for delegated
metadata (add_key()/remove_key()) but the other delegator Targets does
not.
It should provide the same/similar functionality.
Signed-off-by: Martin Vrachev <mvrachev@vmware.com>
