~bzr-pqm/bzr/bzr.dev

Case Insensitive File Systems

=============================

Bazaar must be portable across operating-systems and file-systems.  While the

primary file-system for an operating-system might have some particular

characteristics, it's not necessary that *all* file-systems for that

operating-system will have the same characteristics.

For example, the FAT32 file-system is most commonly found on Windows operating

systems, and has the characteristics usually associated with a Windows

file-system.  However, USB devices means FAT32 file-systems are often used

with Linux, so the current operating system doesn't necessarily reflect the

capabilities of the file-system.

Bazaar supports 3 kinds of file-systems, each to different degrees.

* Case-sensitive file-systems: This is the file-system generally used on

  Linux - 2 files can differ only by case, and the exact case must be used

  when opening a file.

* Case-insensitive, case-preserving (cicp) file-systems: This is the

  file-system generally used on Windows; FAT32 is an example of such a

  file-system.  Although existing files can be opened using any case, the

  exact case used to create the file is preserved and available for programs

  to query.  Two files that differ only by case is not allowed.

* Case-insensitive: This is the file-system used by very old Windows versions

  and is rarely encountered "in the wild".  Two files that differ only by

  case is not allowed and the case used to create a file is not preserved.

As can be implied by the above descriptions, only the first two are considered

relevant to a modern Bazaar.

For more details, including use cases, please see

http://bazaar-vcs.org/CasePreservingWorkingTreeUseCases

Handling these file-systems

---------------------------

The fundamental problem handling these file-systems is that the user may

specify a file name or inventory item with an "incorrect" case - where

"incorrect" simply means different than what is stored - from the user's

point-of-view, the filename is still correct, as it can be used to open, edit

delete etc the item.

The approach Bazaar takes is to "fixup" each of the command-line arguments

which refer to a filename or an inventory item - where "fixup" means to

adjust the case specified by the user so it exactly matches an existing item.

There are two places this match can be performed against - the file-system

and the Bazaar inventory.  When looking at a case-insensitive file-system, it

is impossible to have 2 names that differ only by case, so there is no

ambiguity. The inventory doesn't have the same rules, but it is expected that

projects which wish to work with Windows would, by convention, avoid filenames

that differ only by case.

The rules for such fixups turn out to be quite simple:

* If an argument refers to an existing inventory item, we fixup the argument

  using the inventory.  This is, basically, all commands that take a filename

  or directory argument *other* than 'add' and in some cases 'mv'

* If an argument refers to an existing filename for the creation of an

  inventory item (eg, add), then the case of the existing file on the disk

  will be used.  However, Bazaar must still check the inventory to prevent

  accidentally creating 2 inventory items that differ only by case.

* If an argument results in the creation of a *new* filename (eg, a move

  destination), the argument will be used as specified.  Bzr will create

  a file and inventory item that exactly matches the case specified (although

  as above, care must be taken to avoid creating two inventory items that

  differ only by case.)

Implementation of support for these file-systems

------------------------------------------------

From the description above, it can be seen the implementation is fairly

simple and need not intrude on the internals of Bazaar too much; most of

the time it is simply converting a string specified by the user to the

"canonical" form as stored in either the inventory or filesystem.  These

boil down to the following new API functions:

* osutils.canonical_relpath() - like osutils.relpath() but adjust the case

  of the result to match any existing items.

* Tree.get_canonical_inventory_path - somewhat like Tree.get_symlink_target(),

  Tree.get_file_by_path() etc; returns a name with the case adjusted to match

  existing inventory items.

* osutils.canonical_relpaths() and Tree.get_canonical_inventory_paths() - like

  the 'singular' versions above, but accept and return sequences and therefore

  offer more optimization opportunities when working with multiple names.

The only complication is the requirement that Bazaar not allow the creation

of items that differ only by case on such file-systems.  For this requirement,

case-insensitive and cicp file-systems can be treated the same.  The

'case_sensitive' attribute on a MutableTree is used to control this behaviour.