Content filtering is the feature by which Bazaar can do line-ending conversion or keyword expansion so that the files that appear in the working tree are not precisely the same as the files stored in the repository.
This document describes the implementation; see the user guide for how to use it.
We distinguish between the canonical form which is stored in the repository and the convenient form which is stored in the working tree. The convenient form will for example use OS-local newline conventions or have keywords expanded, and the canonical form will not. We use these names rather than eg "filtered" and "unfiltered" because filters are applied when both reading and writing so those names might cause confusion.
Content filtering is only active on working trees that support it, which is format 2a and later.
Content filtering is configured by rules that match file patterns.
Filters come in pairs: a read filter (reading convenient->canonical) and a write filter. There is no requirement that they be symmetric or that they be deterministic from the input, though in general both these properties will be true. Filters are allowed to change the size of the content, and things like line-ending conversion commonly will.
Filters are fed a sequence of byte chunks (so that they don't have to hold the whole file in memory). There is no guarantee that the chunks will be aligned with line endings. Write filters are passed a context object through which they can obtain some information about eg which file they're working on. (See bzrlib.filters docstring.)
These are at the moment strictly content filters: they can't make changes to the tree like changing the execute bit, file types, or adding/removing entries.
bzrlib interfaces that aren't explicitly specified to deal with the convenient form should return the canonical form. Whenever we have the SHA1 hash of a file, it's the hash of the canonical form.
The dirstate file should store, in the column for the working copy, the cached hash and size of the canonical form, and the packed stat fingerprint for which that cache is valid. This implies that the stored size will in general be different to the size in the packed stat. (However, it may not always do this correctly - see <https://bugs.edge.launchpad.net/bzr/+bug/418439>.)
The dirstate is given a SHA1Provider instance by its tree. This class can calculate the (canonical) hash and size given a filename. This provides a hook by which the working tree can make sure that when the dirstate needs to get the hash of the file, it takes the filters into account.
Most commands that deal with the text of files present the canonical form. Some have options to choose.
Content filters can have serious performance implications. For example, getting the size of (the canonical form of) a file is easy and fast when there are no content filters: we simply stat it. However, when there are filters that might change the size of the file, determining the length of the canonical form requires reading in and filtering the whole file.
Formats from 1.14 onwards support content filtering, so having fast paths for the case where content filtering is not possible is not generally worthwhile. In fact, they're probably harmful by causing extra edges in test coverage and performance.
We need to have things be fast even when filters are in use and then possibly do a bit less work when there are no filters configured.
We might benefit from having filters declare some of their properties statically, for example that they're deterministic or can round-trip or won't change the length of the file. However, common cases like crlf conversion are not guaranteed to round-trip and may change the length, so perhaps adding separate cases will just complicate the code and tests. So overall this does not seem worthwhile.
In a future workingtree format, it might be better not to separately store the working-copy hash and size, but rather just a stat fingerprint at which point it was known to have the same canonical form as the basis tree.
It may be worthwhile to have a virtual Tree-like object that does filtering, so there's a clean separation of filtering from the on-disk state and the meaning of any object is clear. This would have some risk of bugs where either code holds the wrong object, or their state becomes inconsistent.
This would be useful in allowing you to get a filtered view of a historical tree, eg to export it or diff it. At the moment export needs to have its own code to do the filtering.
The convenient-form tree would talk to disk, and the convenient-form tree would sit on top of that and be used by most other bzr code.
If we do this, we'd need to handle the fact that the on-disk tree, which generally deals with all of the IO and generally works entirely in convenient form, would also need to be told the canonical hash to store in the dirstate. This can perhaps be handled by the SHA1Provider or a similar hook.
Content filtering at the moment is a bit specific to on-disk trees: for instance SHA1Provider goes directly to disk, but it seems like this is not necessary.