| Age | Commit message (Collapse) | Author | Lines |
|
|
|
|
|
|
|
The readme is now escaped plain text unless `about-filter` points at
the new `about-render.lua`, which renders markdown, man pages and plain
text server-side. `enable-markdown` goes away with the renderer.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
The diff family inflated a blob of any size just to render it. A file
over the limit is reported as binary instead.
|
|
A dangling or unreadable ref was described through a null object.
|
|
|
|
|
|
|
|
Apple clang still defaults to `-fcommon`, so git's 64 KB packet_buffer
became a common symbol whose size drove a 32 KB section alignment that
ld64 then clamped to the 16 KB page size, warning on every macOS link.
Everywhere else already compiles with `-fno-common`.
|
|
`max-stats` only bounds the selectable periods now and no longer
doubles as the enable switch.
The tree walk runs before the history walk on purpose. Releasing
commit memory while walking history resets each commit slab index,
and a commit graph lookup afterwards would read another commit slot
and walk the wrong tree. The stats fixture writes a commit graph so
the tests cover that path. The history walk bounds the window in
process rather than passing a formatted since date to
`setup_revisions`, and parses each commit once.
|
|
A checkout without tags made `git describe` fail and the empty result
overwrote the fallback, so CI release binaries carried no version at
all. The release workflow now also fetches tags so a tag build stamps
its exact version.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
AUTHORS now records the upstream developers as the project's history
rather than as maintainers of the fork.
|
|
|
|
Clone urls and atom links are built from the request scheme and Host,
but the cache key left them out, so a request with a spoofed Host
could cache a page carrying a bogus clone url and serve it to other
visitors.
|
|
The stat-only diff path returned before the closing call, so the
content wrapper, footer and closing tags were never emitted.
|
|
The feed-level element was written only inside the commit loop, so a
feed with no commits omitted an element the atom format requires.
|
|
The diffstat divided by zero when a commit changed no lines, and a
submodule change was printed with a hardcoded length that dropped a
digit of a sha-1 and most of a sha-256.
|
|
The pager html-escaped its search and sort values instead of
url-encoding them, so a term with a hash or ampersand broke the link
and dropped the active filter on the next page.
|
|
The POST length was clamped with a signed comparison, so a very large
Content-Length could turn negative and slip past the limit into the
fixed-size buffer.
|
|
The idle-sort comparator returned a 64-bit time difference truncated
to int, which could flip sign and leave the ordering inconsistent.
|
|
The index derived a repository's age from `refs/heads/master` alone,
so a repository on any other default branch showed no age. HEAD is
repo-controlled and the age stat walks under `refs/heads`, so a branch
name carrying a parent-directory component is rejected.
|
|
The default sorted the sections and the repositories within them by
name, so the order written in the cgitrc file was ignored.
|
|
cgit unsets HOME to isolate git from the calling user, but git still
finds the global config through its getpwuid fallback and then dies
expanding a `~` in `core.excludesfile`. That die happened after the
snapshot headers and the gzip filter were already in place, so the
error page was compressed into the archive and every snapshot came
out undecompressable.
|
|
The log page passed the `id` query parameter to git's revision parser
without resolving it and ahead of the end-of-options marker. A value
like `id=--output=/path` was then parsed as an option, so an
unauthenticated request could create or truncate any file the server
user could write. No valid ref or object name begins with a dash.
|
|
A crafted `ofs` could send cgit walking most of the history for one
request.
|
|
The about subpath was confined to the readme directory with a plain
byte-prefix match, so a sibling directory sharing the base name as a
prefix passed the check and its files were served.
|
|
The line-number gutter did one allocation and one write per line.
|
|
Only the tree view honoured the limit, and only after loading the
whole object. The raw blob, plain, blame and readme paths now check
the size before reading, and the default moves from unlimited to
10 MB so a fresh install never buffers a huge object whole, with zero
still the opt-out.
|
|
A blob requested by ref with an unknown path fell through to the
commit object and was served with a 200 instead of a 404, two error
pages passed a null pointer to a %s format when the request carried
no object id, and the error pages left the layout open.
|
|
`fmt()` aborts the request rather than truncate, so a very long path
fed into a context-sensitive tab title took down page rendering.
|
|
The year period rendered without its label, and a commit with no
author line crashed the view.
|
|
A plain char is undefined in the ctype functions for negative values,
which a high byte produces wherever char is signed.
|