Don’t really need the hashes of the current versions - just knowing whether they’ve changed or not will generally be enough - and just the mtime and ctime of a point in time may be enough?


There are currently 4 levels that state can have.

  1. NOT_IN_MEMORY The actual content blocks have not been read at all.
  2. IN_MEMORY_UNMODIFIED The content blocks have been read and are available for use. They have not been changed at all versus what was written on disk when we read them.
  3. IN_MEMORY_HASH_MODIFIED We have updated the in-memory state, but only to record the sha1/symlink target value and the stat value that means this information is ‘fresh’.
  4. IN_MEMORY_MODIFIED We have updated an actual record. (Parent lists, added a new file, deleted something, etc.) In this state, we must always write out the dirstate, or some user action will be lost.


This state is a bit special, so deserves its own topic. If we are IN_MEMORY_HASH_MODIFIED, we only write out the dirstate if enough records have been updated. The idea is that if we would save future I/O by writing an updated dirstate, then we should do so. The threshold for this is set by “worth_saving_limit”. The default is that at least 10 entries must be updated in order to consider the dirstate file worth updating.

Going one step further, newly added files, symlinks, and directory entries updates are treated specially. We know that we will always stat all entries in the tree so that we can observe if they have changed. In the case of directories, all the information we know about them is just from that stat value. There is no extra content to read. So an update directory entry doesn’t cause us to update to IN_MEMORY_HASH_MODIFIED. However, if there are other modifications worth saving, we will go ahead and save the directory entry update at the same time.

Similarly, symlink targets are commonly stored in the inode entry directly. So once we have stat’ed the symlink, we already have its target information in memory. The one caveat is if we used to think an object was a file, and it became a directory or symlink, then we will treat it as worth saving.

In the case of newly added files, we never have to read their content to know that they are different from the basis tree. So saving the updated information also won’t save a future read.

Table Of Contents

Previous topic

Miscellaneous notes

Next topic

“bzr update” performance analysis

This Page