| Age | Commit message (Collapse) | Author | Lines |
|
A repos file the filter could not open read as listing nothing, so a
permission slip or a mistyped path made every protected repository
public. A line naming a repository with no groups was skipped the same
way, where auth-inline's empty list denies everyone.
|
|
Five copies of the test and four of the error page had to agree by
hand. The new check pins the four refusals, which no test covered.
|
|
A failed stat logged the directory rather than the entry inside it,
and a project list that failed to read printed strerror of the flag
ferror returns, which names an unrelated error.
|
|
Matching xml anywhere in the type also caught the Office formats, whose
names hold openxmlformats, so a .docx went out with a UTF-8 charset.
|
|
Only a hash in the first column marked a comment, so an indented one
was read as the type "#" for every word after it. The manual already
describes comments by their first non-blank character.
|
|
The dynamic ttl only applied when the url named a branch with h=, so
the log, tree and refs of the default branch, which never carry one,
fell back to cache-summary-ttl. The manual already gives each ttl the
scope the code now uses.
|
|
The guard took local as a name that may carry the -local suffix, but
git reads local only as a whole alias of default-local and dies on
anything after it, so that one value turned every page into a 500.
|
|
The commit and diff pages collect each file's body before printing it,
and a die in git while one was collected wrote the error page into that
body, which was never sent, so the visitor got a page cut off mid table.
|
|
trust-scan-config passed any readme holding a colon as a tracked file,
while the parser takes one only when a path follows the colon, so a
scanned repository could still name a file on disk outside itself. The
two now share one test.
|
|
The field resets at the end of cgit_add_repo cleared the ignore flag
that the empty url branch had just set, so such a repository was
listed on the index and selectable with ?r=, where every link to it
read the byte before its url.
|
|
The generator meta, the footer and the patch signature named the exact
cgit and git versions on every response, which only helps someone
matching a site against an advisory. The shipped server configs
already hide the server's own version for the same reason, and
`cgit --version` still answers on the host.
|
|
Only the login page carried a Cache-Control, so a page an auth filter
had let a visitor see could be kept by a cache shared with the next
visitor. The page also varies on the cookie that got them in.
|
|
Each class behind cgit's published advisories now has a check against
the current code, from a newline in a file name to shell syntax handed
to a filter. The feed also treated an ident with an empty name and an
empty address as present and wrote an empty person, which Atom forbids.
|
|
A write that failed with EPIPE on an uncached page died into an error
page, which died again at exit when the page could not be sent either,
and git logged a recursion warning on every dropped connection. A
filter that exits early fails a write the same way, so the quiet exit
is kept for writes that reach the client itself.
|
|
Archives, packs and the HEAD file went out as octet-stream or gzip
with a UTF-8 charset on the end, since only the blob page cleared it.
The headers now decide from the type itself, so the two places that
cleared the charset by hand no longer need to.
|
|
Such a name is refused as a head before git could read it as an
option, and the fallback branch already skips it, so the switcher
offered a choice that led only to an error page.
|
|
A commit whose encoding header names something iconv cannot convert,
or whose text is labelled UTF-8 without being it, reached a page
declared UTF-8 with its bytes untouched. The atom feed already
replaced them, so its UTF-8 reader moves to shared.c and every ident,
subject, message and tag text passes through it.
|
|
With both an id and a path the blob page handed back the raw object
the id named, so a commit id gave the commit text itself, while the
plain page looked the path up in that commit's tree. A tag is peeled
to its commit on the way, and the README example of a pinned commit
now uses a form the request checks accept.
|
|
The hop to the trailing-slash form of the about page, and the hop back
to the summary of a repository without a readme, were built from the
path alone, so a request for the about page of another branch landed
on the default one. The query goes into the Location line as the
client sent it, with any byte a header cannot carry percent-encoded.
|
|
It lists object directories by their path on the server, which a
client fetching over http can neither use nor needs to learn. The
http-alternates file written for such clients still goes out.
|
|
The diff and patch pages honour the path in the url, which the
parameter table documents for every page, but rawdiff never handed it
to git and answered with the whole diff.
|
|
The dumb transport read a pack four kilobytes at a time through stdio
and sent it without a length, so a client could not tell a cut-off
transfer from a complete one.
|
|
The dumb transport reads files that already sit on the disk, so a pack
copied into a slot cost that disk twice and the request a second write
of every byte. A snapshot took a slot whatever its size, so a visitor
naming distinct refs and ids could fill the cache root with archives.
`cache-max-slot-size`, 64 MB unless set, now serves a larger response
from the lock file and drops it, along with any expired copy it would
have replaced.
|
|
`enable-html-serving` makes the plain page send a repository file as
text/html on the site's own origin, with no nosniff and no policy, so
a scanned repository could switch it on from its git config or cgitrc
without `trust-scan-config` and run script against every visitor. The
warning for a key read from git config also named a null repository,
because `repo->path` was set only after that file had been read.
|
|
`cgit_abort_filters` unhooks the Lua write interposer under NO_LUA
too, where neither the hook nor its state exists, so `make NO_LUA=1`
and the default mode of `tools/release-build.sh` have not compiled
since the die path learned to take stdout back from a filter. The
suite now builds that variant into `build/nolua` and runs it, and a
`lua:` filter in such a build is refused with a message naming the
cause instead of an unknown filter type.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
`compose_snapshot_prefix` dropped the leading v of a tag only when just
one of the names 1.2, v1.2 and V1.2 resolved as a tag, so that the
shorter snapshot name could always be traced back to one tag. Those
probes went through ref lookups, and on a case-insensitive filesystem
a lookup for V1.2 finds the loose file of v1.2, so every freshly made
tag counted as ambiguous and kept its v. Once git packed the refs the
lookups became exact and the same tag quietly changed its snapshot
names.
The claimants on a stripped name are now counted over the tag list
itself with exact string comparison, so the answer no longer depends
on how a ref is stored or on the filesystem underneath. Two tags that
really differ only by the letter's case still both keep it.
|
|
|
|
|
|
The inline handlers and the auto-submitting selects are gone, so
`script-src` no longer needs it, and t0004 now checks that the three
configs pin the same policy.
|
|
|
|
|
|
|
|
|
|
|
|
A `max-stats` period enables the page again, as it did before
v2.2.0, so one key does both jobs.
|
|
|
|
|