[PATCH jn/svn-fe-maint 0/2] intermittent repo_tree corruption

[Jonathan Nieder <jrnieder@gmail.com>]

[mailer usage fail.  resending to git@vger; sorry for the noise]

Hi David,

Jonathan Nieder wrote:
> Jonathan Nieder wrote:

>> +++ b/vcs-svn/svndump.c
>> @@ -181,12 +181,22 @@ static void handle_node(void)
> [...]
>> +		if (mode == REPO_MODE_DIR && type != REPO_MODE_DIR)
>> +			die("invalid dump: cannot modify a directory into a file");
>
> This is tripping for me in a revision like so:
[...] 
>  Node-path: lxanew/DCNsnetwork.cpp
>  Node-kind: file
>  Node-action: add
[...]
> DCNsnetwork.cpp was added in r688, which looks like so:
[...]
>  Node-path: DCNsnetwork.cpp
>  Node-kind: file
>  Node-action: add

This is a bad one.  Currently, repo_copy builds new nodes for all
directories containing the destination directory, like so:

	for each path component:
		find that node, creating if missing
		make sure it is in the active commit (cloning if not)
		set mode to MODE_DIR and set up the next iteration to look there.
	set mode and content based on source path

Unfortunately the "create if missing" and "make sure it is the active
commit" steps call realloc without updating the corresponding
pointers.

svn-fe2 (as contained in git 1.7.3) and svn-fe3 are both affected,
presumably.  Probably existing testing missed this because the
repo_tree pool quickly grows large enough that glibc allocates with
mmap(), with little need to change addresses within a huge 64-bit
address space.

Jonathan Nieder (2):
  treap: make treap_insert return inserted node
  vcs-svn: fix intermittent repo_tree corruption

 test-treap.c        |   11 ++++++++---
 vcs-svn/repo_tree.c |    4 ++--
 vcs-svn/trp.h       |    3 ++-
 vcs-svn/trp.txt     |   10 ++++++++--
 4 files changed, 20 insertions(+), 8 deletions(-)

-- 
1.7.2.4