Maybe it already got one of the corrupted hashes from a different server and it doesn't double-check that the hash from every server is correct. Or another problem. But in any case I'm marking this as TODO because an even better (more picky) verifier is less urgent than repairer.
New checker and verifier use the new download class. They are robust against various sorts of failures or corruption. They return detailed results explaining what they learned about your immutable files. Some grotesque sorts of corruption are not properly handled yet, and those ones are marked as TODO or commented-out in the unit tests.
There is also a repairer module in this patch with the beginnings of a repairer in it. That repairer is mostly just the interface to the outside world -- the core operation of actually reconstructing the missing data blocks and uploading them is not in there yet.
This patch also refactors the unit tests in test_immutable so that the handling of each kind of corruption is reported as passing or failing separately, can be separately TODO'ified, etc. The unit tests are also improved in various ways to require more of the code under test or to stop requiring unreasonable things of it. :-)
The code for validating the share hash tree and the block hash tree has been rewritten to make sure it handles all cases, to share metadata about the file (such as the share hash tree, block hash trees, and UEB) among different share downloads, and not to require hashes to be stored on the server unnecessarily, such as the roots of the block hash trees (not needed since they are also the leaves of the share hash tree), and the root of the share hash tree (not needed since it is also included in the UEB). It also passes the latest tests including handling corrupted shares well.
ValidatedReadBucketProxy takes a share_hash_tree argument to its constructor, which is a reference to a share hash tree shared by all ValidatedReadBucketProxies for that immutable file download.
ValidatedReadBucketProxy requires the block_size and share_size to be provided in its constructor, and it then uses those to compute the offsets and lengths of blocks when it needs them, instead of reading those values out of the share. The user of ValidatedReadBucketProxy therefore has to have first used a ValidatedExtendedURIProxy to compute those two values from the validated contents of the URI. This is pleasingly simplifies safety analysis: the client knows which span of bytes corresponds to a given block from the validated URI data, rather than from the unvalidated data stored on the storage server. It also simplifies unit testing of verifier/repairer, because now it doesn't care about the contents of the "share size" and "block size" fields in the share. It does not relieve the need for share data v2 layout, because we still need to store and retrieve the offsets of the fields which come after the share data, therefore we still need to use share data v2 with its 8-byte fields if we want to store share data larger than about 2^32.
Specify which subset of the block hashes and share hashes you need while downloading a particular share. In the future this will hopefully be used to fetch only a subset, for network efficiency, but currently all of them are fetched, regardless of which subset you specify.
ReadBucketProxy hides the question of whether it has "started" or not (sent a request to the server to get metadata) from its user.
Download is optimized to do as few roundtrips and as few requests as possible, hopefully speeding up download a bit.
This does raise the question of if there is any point to this test, since I apparently don't know what the answer *should* be, and whenever one of the buildbots fails then I redefine success.
But, I'm about to commit a bunch of patches to implement checker, verifier, and repairer as well as to refactor downloader, and I would really like to know if these patches *increase* the number of reads required even higher than it currently is.
Nathan Wilcox observed that the storage server can rely on the size of the share file combined with the count of leases to unambiguously identify the location of the leases. This means that it can hold any size share data, even though the field nominally used to hold the size of the share data is only 32 bits wide.
With this patch, the storage server still writes the "size of the share data" field (just in case the server gets downgraded to an earlier version which requires that field, or the share file gets moved to another server which is of an earlier vintage), but it doesn't use it. Also, with this patch, the server no longer rejects requests to write shares which are >= 2^32 bytes in size, and it no longer rejects attempts to read such shares.
This fixes http://allmydata.org/trac/tahoe/ticket/346 (increase share-size field to 8 bytes, remove 12GiB filesize limit), although there remains open a question of how clients know that a given server can handle large shares (by using the new versioning scheme, probably).
Note that share size is also limited by another factor -- how big of a file we can store on the local filesystem on the server. Currently allmydata.com typically uses ext3 and I think we typically have block size = 4 KiB, which means that the largest file is about 2 TiB. Also, the hard drives themselves are only 1 TB, so the largest share is definitely slightly less than 1 TB, which means (when K == 3), the largest file is less than 3 TB.
This patch also refactors the creation of new sharefiles so that only a single fopen() is used.
This patch also helps with the unit-testing of repairer, since formerly it was unclear what repairer should expect to find if the "share data size" field was corrupted (some corruptions would have no effect, others would cause failure to download). Now it is clear that repairer is not required to notice if this field is corrupted since it has no effect on download. :-)
There are a lot of different ways that a share could be corrupted, or that attempting to download it might fail. These tests attempt to exercise many of those ways and require the checker/verifier/repairer to handle each kind of failure well.
In an ancient version of directories, we needed a MAC on each entry. In modern times, the entire dirnode comes with a digital signature, so the MAC on each entry is redundant.
With this patch, we no longer check those MACs when reading directories, but we still produce them so that older readers will accept directories that we write.
I get confused about whether a given argument or return value is a uri-as-string or uri-as-object. This patch adds a lot of assertions that it is one or the other, and also changes CheckerResults to take objects not strings.
In the future, I hope that we generally use Python objects except when importing into or exporting from the Python interpreter e.g. over the wire, the UI, or a stored file.
Tahoe webapi was failing on HTTP request containing a partial Range header.
This change allows movies players like mplayer to seek in movie files stored in
tahoe.
Associated tests for GET and HEAD methods are also included
Refactor into a class the logic of asking each server in turn until one of them gives an answer
that validates. It is called ValidatedThingObtainer.
Refactor the downloading and verification of the URI Extension Block into a class named
ValidatedExtendedURIProxy.
The new logic of validating UEBs is minimalist: it doesn't require the UEB to contain any
unncessary information, but of course it still accepts such information for backwards
compatibility (so that this new download code is able to download files uploaded with old, and
for that matter with current, upload code).
The new logic of validating UEBs follows the practice of doing all validation up front. This
practice advises one to isolate the validation of incoming data into one place, so that all of
the rest of the code can assume only valid data.
If any redundant information is present in the UEB+URI, the new code cross-checks and asserts
that it is all fully consistent. This closes some issues where the uploader could have
uploaded inconsistent redundant data, which would probably have caused the old downloader to
simply reject that download after getting a Python exception, but perhaps could have caused
greater harm to the old downloader.
I removed the notion of selecting an erasure codec from codec.py based on the string that was
passed in the UEB. Currently "crs" is the only such string that works, so
"_assert(codec_name == 'crs')" is simpler and more explicit. This is also in keeping with the
"validate up front" strategy -- now if someone sets a different string than "crs" in their UEB,
the downloader will reject the download in the "validate this UEB" function instead of in a
separate "select the codec instance" function.
I removed the code to check plaintext hashes and plaintext Merkle Trees. Uploaders do not
produce this information any more (since it potentially exposes confidential information about
the file), and the unit tests for it were disabled. The downloader before this patch would
check that plaintext hash or plaintext merkle tree if they were present, but not complain if
they were absent. The new downloader in this patch complains if they are present and doesn't
check them. (We might in the future re-introduce such hashes over the plaintext, but encrypt
the hashes which are stored in the UEB to preserve confidentiality. This would be a double-
check on the correctness of our own source code -- the current Merkle Tree over the ciphertext
is already sufficient to guarantee the integrity of the download unless there is a bug in our
Merkle Tree or AES implementation.)
This patch increases the lines-of-code count by 8 (from 17,770 to 17,778), and reduces the
uncovered-by-tests lines-of-code count by 24 (from 1408 to 1384). Those numbers would be more
meaningful if we omitted src/allmydata/util/ from the test-coverage statistics.
