Philip Martin <phi...@codematters.co.uk> writes:
> My reproduction doesn't trigger the bug in 1.8 but that seems to be
> because 1.8 has some other problem and the rep-cache deduplication
> doesn't trigger. The third commit restoring the content in the first
> commit simply stores a new delta.
Comparing 1.8 and 1.9 the reason that 1.8 doesn't do deduplication is
this code, fs_fs.c:get_shared_rep:7757
/* We don't want 0-length PLAIN representations to replace non-0-length
ones (see issue #4554). Also, this doubles as a simple guard against
general rep-cache induced corruption. */
if ( ((*old_rep)->expanded_size != rep->expanded_size)
|| ((*old_rep)->size != rep->size))
{
*old_rep = NULL;
}
else
In 1.9 the corresponding code is transaction.c:get_shared_rep:2234
if ( ((*old_rep)->expanded_size != rep->expanded_size)
|| ((rep->expanded_size == 0) && ((*old_rep)->size != rep->size)))
{
*old_rep = NULL;
}
else
Since 1.8 doesn't have the expanded_size check it sets *old_rep to NULL
and therefore doesn't do deduplication in some cases where it would
otherwise be possible. If I patch 1.8
Index: subversion/libsvn_fs_fs/fs_fs.c
===================================================================
--- subversion/libsvn_fs_fs/fs_fs.c (revision 1826211)
+++ subversion/libsvn_fs_fs/fs_fs.c (working copy)
@@ -7754,7 +7754,7 @@ get_shared_rep(representation_t **old_rep,
ones (see issue #4554). Also, this doubles as a simple guard against
general rep-cache induced corruption. */
if ( ((*old_rep)->expanded_size != rep->expanded_size)
- || ((*old_rep)->size != rep->size))
+ || ((rep->expanded_size == 0) && (*old_rep)->size != rep->size))
{
*old_rep = NULL;
}
the deduplication happens (and the checksum failure does not occur
because the checksum bug does not affect 1.8).
--
Philip
