Sparse complains that we are using a plain integer as a NULL pointer
here, but in fact we do not have to specify a value for this variable at
all since it has static storage duration and thus will be initialized to
NULL by the compiler.
Signed-off-by: John Keeping <john@keeping.me.uk>
If CGit is killed while it holds a lock on a cache slot (for example
because it is taking too long to generate a page), the lock file will be
left in place. This prevents any future attempt to use the same slot
since it will fail to exclusively create the lock file.
Since CGit is the only program that should be manipulating lock files,
we can use advisory locking to detect whether another process is
actually using the lock file or if it is now stale.
I have confirmed that this works on Linux by setting a short TTL in a
custom cgitrc and running the following with CGit patched to print a
message to stderr if the fcntl(2) fails:
$ export CGIT_CONFIG=$PWD/cgitrc
$ export QUERY_STRING=url=cgit/tree/ui-shared.c
$ ./cgit |
grep -v -e '^<div class=.footer.>' \
-e '^Last-Modified: ' \
-e ^'Expires: ' >expect
$ seq 50000 | dd bs=8192 |
parallel -j200 "diff -u expect <(./cgit |
grep -v -e '^<div class=.footer.>' \
-e '^Last-Modified: ' \
-e ^'Expires: ') || echo BAD"
This printed the fail message several times without ever printing "BAD".
Signed-off-by: John Keeping <john@keeping.me.uk>
When clicking on "log" from a tag we end up showing the log of whatever
branch we used to reach the tag. If the tag doesn't point onto a branch
then the tagged commit won't appear in this output.
By linking to tags with the head parameter instead of the "id" parameter
the log link will show the log of the tag. This is clearly desirable
when the tag has been reached from the refs UI and changing the
behaviour for tag decorations makes them match branch decorations where
log -> decoration -> log shows the log of the decoration.
Reported-by: Ferry Huberts <mailings@hupie.com>
Signed-off-by: John Keeping <john@keeping.me.uk>
Seeing the diff stat for a single file is pretty useless, so reset the
diff type before generating the links to individual files in the diff
stat so that the links will show a useful diff.
Reported-by: Konstantin Ryabitsev <mricon@kernel.org>
Signed-off-by: John Keeping <john@keeping.me.uk>
Using (DIFF_FORMAT_DIFFSTAT | DIFF_FORMAT_PATCH) causes Git to emit a
"---" line between the commit message and the body of the patch, which
fixes a regression introduced in commit 455b598 (ui-patch.c: Use
log_tree_commit() to generate diffs, 2013-08-20), prior to which we
inserted the "---" line ourselves.
DIFF_FORMAT_SUMMARY is added so that we match the output of
git-format-patch(1) without the "-p" option.
Signed-off-by: John Keeping <john@keeping.me.uk>
* &&-chaining
* use test_cmp instead of cmp
* use strip_headers instead of knowing how many lines there will be
Signed-off-by: John Keeping <john@keeping.me.uk>
This will allow us to reuse the same logic to add clone URL <link/>
elements to the header of all repo-specific pages in order to support
the rel-vcs microformat.
Signed-off-by: John Keeping <john@keeping.me.uk>
This prints the diffstat but stops before printing (or generating) any
of the body of the diff.
No cgitrc option is added here so that we can wait to see how useful
this is before letting people set it as the default.
Suggested-by: Konstantin Ryabitsev <mricon@kernel.org>
Signed-off-by: John Keeping <john@keeping.me.uk>
This will allow us to introduce a new "stat only" diff mode without
needing an explosion of mutually incompatible flags.
The old "ss" query parameter is still accepted in order to avoid
breaking saved links, but we no longer generate any URIs using it;
instead the new "dt" (diff type) parameter is used.
Signed-off-by: John Keeping <john@keeping.me.uk>
This is slightly more involved than just bumping the version number
because it pulls in a change to convert the commit buffer to a slab,
removing the "buffer" field from "struct commit". All sites that access
"commit->buffer" have been changed to use the new functions provided for
this purpose.
Signed-off-by: John Keeping <john@keeping.me.uk>
It looks like cached patches are truncated to the nearest 1024-byte
boundary in the patch body. E.g.:
> mricon@nikko:[/tmp]$ wget -O no-cache
> "http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/patch/?id=6e1b4fdad5157bb9e88777d525704aba24389bee"
...
> 2014-06-11 15:34:51 (80.4 MB/s) - ‘no-cache’ saved [4767]
Patch is complete, without truncation. Next hit, with cache in place:
> mricon@nikko:[/tmp]$ wget -O yes-cache
> "http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/patch/?id=6e1b4
> fdad5157bb9e88777d525704aba24389bee"
...
> 2014-06-11 15:35:01 (17.0 MB/s) - ‘yes-cache’ saved [4096/4096]
Length truncated to 4096. The cache on disk looks truncated as well, so
the bug must me during the process of saving cache. The same is true for
larger patches:
> mricon@nikko:[/tmp]$ wget -O no-cache
> "http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/patch/?id=2840c566e95599cd60c7143762ca8b49d9395050"
...
> 2014-06-11 15:41:33 (1.07 MB/s) - ‘no-cache’ saved [979644]
979644 bytes with a cache-miss
> mricon@nikko:[/tmp]$ wget -O yes-cache
> "http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/patch/?id=2840c
> 566e95599cd60c7143762ca8b49d9395050"
...
> 2014-06-11 15:41:46 (1.05 MB/s) - ‘yes-cache’ saved [978944]
978944 (956KB exactly) with a cache-hit
Since the "html" functions use raw write(2) to STDIO_FILENO, we don't
notice problems with most pages, but raw patches write using printf(3).
This is fine if we're outputting straight to stdout since the buffers
are flushed on exit, but we close the cache output before this, so the
cached output ends up being truncated.
Make sure the buffers are flushed when we finish outputting a patch so
that we avoid this.
No other UIs use printf(3) so we do not need to worry about them.
Actually, it's slightly more interesting than this... since we don't set
GIT_FLUSH, Git decides whether or not it will flush stdout after writing
each commit based on whether or not stdout points to a regular file (in
maybe_flush_or_die()).
Which means that when writing directly to the webserver, Git flushes
stdout for us, but when we redirect stdout to the cache it points to a
regular file so Git no longer flushes the output for us.
The patch is still correct, but perhaps the full explanation is
interesting!
Reported-by: Konstantin Ryabitsev <mricon@kernel.org>
If you search for a bogus range string here:
http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/log/
Using something like "range" and "qwerty123456", it returns an "Internal
Server Error" and the following in the logs:
> [Tue Jun 10 17:45:32 2014] [error] [client 172.21.1.6] fatal:
> ambiguous argument 'qwerty123456': unknown revision or path not in the
> working tree., referer:
> http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/
> [Tue Jun 10 17:45:32 2014] [error] [client 172.21.1.6] Use '--' to
> separate paths from revisions, like this:, referer:
> http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/
> [Tue Jun 10 17:45:32 2014] [error] [client 172.21.1.6] 'git <command>
> [<revision>...] -- [<file>...]', referer:
> http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/
> [Tue Jun 10 17:45:32 2014] [error] [client 172.21.1.6] Premature end
> of script headers: cgit, referer:
> http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/
The cache will kick in, so if you search for the same string again,
it'll show an empty range, so you have to change the bogus strings each
time.
This is because we just pass the arguments straight to Git's revision
parsing machinery which die()s if it cannot parse an argument, printing
the above to stderr and exiting.
The patch below makes it a bit friendlier by just ignoring unhandled
arguments, but I can't see an easy way to report errors when we can't
parse revision arguments without losing the flexibility of supporting
all of the revision specifiers supported by Git.
Reported-by: Konstantin Ryabitsev <mricon@kernel.org>
This allows different filter implementations to be specified in the
configuration file. Currently only "exec" is supported, but it may now
be specified either with or without the "exec:" prefix.
Signed-off-by: John Keeping <john@keeping.me.uk>
Change the existing cgit_{open,close,fprintf}_filter functions to
delegate to filter-specific implementations accessed via function
pointers on the cgit_filter object.
We treat the "exec" filter type slightly specially here by putting its
structure definition in the header file and providing an "init" function
to set up the function pointers. This is required so that the
ui-snapshot.c code that applies a compression filter can continue to use
the filter interface to do so.
Signed-off-by: John Keeping <john@keeping.me.uk>
This stops the code in cgit.c::print_repo needing to inspect the
cgit_filter structure, meaning that we can abstract out different filter
types that will have different fields that need to be printed.
Signed-off-by: John Keeping <john@keeping.me.uk>
This allows running the entire test suite with a set of command-line
options. For example:
make test CGIT_TEST_OPTS=--valgrind
Signed-off-by: John Keeping <john@keeping.me.uk>
This is for consistency with other callers. The value returned from
cgit_rooturl is not guaranteed to be HTML-safe.
Signed-off-by: John Keeping <john@keeping.me.uk>
As far as I know, there is no requirement that $SCRIPT_NAME contain only
URL-safe characters, so we need to make sure that any special characters
are escaped.
Signed-off-by: John Keeping <john@keeping.me.uk>
Everywhere else we use html_txt to escape any special characters in
these variables. Do so here as well.
Signed-off-by: John Keeping <john@keeping.me.uk>
This avoids poking into the filter data structure at various points in
the code. We rely on the fact that the number of arguments is fixed
based on the filter type (set in cgit_new_filter) and that the call
sites all know which filter type they're using.
Signed-off-by: John Keeping <john@keeping.me.uk>
By switching the assignment of fields in the cgit_filter structure to
use designated initializers, the compiler will initialize all other
fields to their default value. This will be needed when we add the
extra_args field in the next patch.
Signed-off-by: John Keeping <john@keeping.me.uk>
Now running tests with the "--valgrind" option will run cgit under
Valgrind instead of all Git commands.
Signed-off-by: John Keeping <john@keeping.me.uk>
Valgrind says:
==18344== Conditional jump or move depends on uninitialised value(s)
==18344== at 0x406C83: open_slot (cache.c:63)
==18344== by 0x407478: cache_ls (cache.c:403)
==18344== by 0x404C9A: process_request (cgit.c:639)
==18344== by 0x406BD2: fill_slot (cache.c:190)
==18344== by 0x4071A0: cache_process (cache.c:284)
==18344== by 0x404461: main (cgit.c:952)
==18344== Uninitialised value was created by a stack allocation
==18344== at 0x40738B: cache_ls (cache.c:375)
This is caused by the keylen field being used to calculate whether or
not a slot is matched. We never then check the value of this and the
length of data read depends on the key length read from the file so this
isn't dangerous, but it's nice to avoid branching based on uninitialized
data.
Signed-off-by: John Keeping <john@keeping.me.uk>
When outputting the Content-Type HTTP header we print the MIME type and
then append "; charset=<charset>" if the charset variable is non-null.
We don't want a charset when we have selected "application/octet-stream"
or when the user has specified a custom MIME type, since they may have
specified their own charset. To avoid this, make sure we set the page's
charset to NULL in ui-plain before we generate the HTTP headers.
Signed-off-by: John Keeping <john@keeping.me.uk>
Signed-off-by: Lukas Fleischer <cgit@cryptocrack.de>
By using the standard library's printf, cache_ls does not redirect its
output to the cache when we change the process' stdout file descriptor
to point to the cache file. Fix this by using "htmlf" in the same way
that we do for writing HTTP headers.
Signed-off-by: John Keeping <john@keeping.me.uk>
This means that we can avoid hardcoding the number of headers we expect
CGit to generate in test cases and simply remove whatever headers happen
to by there when we are checking body content.
Signed-off-by: John Keeping <john@keeping.me.uk>
If we fail to write HTML output once, there's no point carrying on so
just write a failure message once and die. By using Git's die_errno
function we also let the user know in what way the write failed.
Signed-off-by: John Keeping <john@keeping.me.uk>
This helps projects that have a large number of tags to display them all
using custom CSS.
The default stylesheet has not been updated since what is useful for
projects with a lot of tags is not the same as what is useful for
projects with only a small number of decorations per commit.
Suggested-by: Konstantin Ryabitsev <mricon@kernel.org>
Signed-off-by: John Keeping <john@keeping.me.uk>
When building the "test" target we depend on both cgit and building the
Git tools. By doing this with two targets we end up running make in the
git/ directory twice, concurrently if using parallel make, which causes
us to build more than we need and potentially builds incorrectly if
multi-step build-then-move operations overlap.
Fix this by instead calling back into the makefile so that we alter the
"cgit" target to also build the Git tools.
Signed-off-by: John Keeping <john@keeping.me.uk>
Commit fb3655d (use struct strbuf instead of static buffers, 2013-04-06)
broke the logic in cache.c::cache_ls by failing to set slot->cache_name
before calling open_slot.
While fixing this, also free the strbufs added by that commit once we're
done with them.
Signed-off-by: John Keeping <john@keeping.me.uk>
We try to stick to POSIX shell in the tests but a "function" keyword has
found its way into t0109. Remove it.
This makes the tests work with dash again.
Signed-off-by: John Keeping <john@keeping.me.uk>
When creating the GIT-VERSION-FILE that we use to test that the version
of Git in git/ is the same as in the CGit Makefile, Git applies the
transform "s/-/./g" to the version string. This doesn't affect released
versions but does change RC version numbers such as 1.8.3-rc0.
While CGit should only refer to a released Git version in general, it is
useful to developers who want to test upcoming Git releases if the tests
do work with RCs, so change t0001 to apply the same transform to our
Makefile version before comparing it to the contents of
GIT-VERSION-FILE.
Signed-off-by: John Keeping <john@keeping.me.uk>
Commit fb3655d (use struct strbuf instead of static buffers -
2013-04-06) introduced a regression in the "section-from-path" handling
when the configured value is negative. By changing the "rel" variable
so that it includes a trailing slash, counting slashes from the end of
the string no longer gives the same answer as it did before.
Fix this by ensuring that "rel" does not have a trailing slash.
Reported-by: Julius Plenz <plenz@cis.fu-berlin.de>
Signed-off-by: John Keeping <john@keeping.me.uk>
When testing modifications in Git that affect CGit, it is annoying to
have t0001 failing simply because the Git version has a ".dirty" suffix
when the version of Git there does indeed match that specified in the
CGit makefile. Stop this by stripping the ".dirty" suffix from the
GIT_VERSION variable.
Note that this brings the "Git version" behaviour in line with the
"submodule version" case which does not check if the working tree in
git/ is modified.
Signed-off-by: John Keeping <john@keeping.me.uk>
By default, Git's test suite puts the trash directories and test-results
directory into its own directory, not that containing the tests being
run. This is less convenient for inspecting test failures, so set the
output directory to CGit's tests/ directory instead.
Note that there is currently a bug in Git whereby it will create the
trash directories in our tests/ directory regardless of the value of
TEST_OUTPUT_DIRECTORY, and then fail to remove them once the tests are
done. This change does currently affect the location of the
test-results/ directory though.
Signed-off-by: John Keeping <john@keeping.me.uk>
In order to ensure that we don't access $HOME at some point after
initial startup when rendering a specific view, run the strace test on a
range of different pages.
This ensures that we don't end up reading a configuration later for some
specific view.
Signed-off-by: John Keeping <john@keeping.me.uk>
Without '&&' between operations, we will not detect if strace or cgit
exit with an error status, which would cause a false positive test
status in this case.
Signed-off-by: John Keeping <john@keeping.me.uk>
This allows tests to run in parallel as well as letting us use "prove"
or another TAP harness to run the tests.
Git's test framework requires Git to be fully built before letting any
tests run, so add a new target to the top-level Makefile which builds
all of Git instead of just libgit.a and make the "test" target depend on
that.
Signed-off-by: John Keeping <john@keeping.me.uk>
Use "struct strbuf" from Git to remove the limit on file path length.
Notes on scan-tree:
This is slightly involved since I decided to pass the strbuf into
add_repo() and modify if whenever a new file name is required, which
should avoid any extra allocations within that function. The pattern
there is to append the filename, use it and then reset the buffer to its
original length (retaining a trailing '/').
Notes on ui-snapshot:
Since write_archive modifies the argv array passed to it we
copy the argv_array values into a new array of char* and then free the
original argv_array structure and the new array without worrying about
what the values now look like.
Signed-off-by: John Keeping <john@keeping.me.uk>
After this change there is one remaining call 'fmt("%s", delim)' in
ui-shared.c but is needed as delim is stack allocated and so cannot be
returned from the function.
Signed-off-by: John Keeping <john@keeping.me.uk>
This removes many uses of "fmt" which uses a fixed size static pool of
fixed size buffers. Instead of relying on these, we now pass around
argument lists for as long as possible before using a strbuf to render
content of an arbitrary size.
Signed-off-by: John Keeping <john@keeping.me.uk>
This adds the fmtalloc helper, html_txtf, html_vtxtf, and html_attrf.
These takes a printf style format string like htmlf but escapes the
resulting string. The html_vtxtf variant takes a va_list whereas
html_txtf is variadic.
Signed-off-by: John Keeping <john@keeping.me.uk>
The CGit configuration variable virtual_root is normalized so that it
does not have a trailing '/' character, but it is allowed to be empty
(the empty string and NULL have different meanings here) and there is
code that is insufficiently cautious when checking if it ends in a '/':
if (virtual_root[strlen(virtual_root) - 1] != '/')
Clearly this check is redundant, but rather than simply removing it we
get a slight efficiency improvement by switching the normalization so
that the virtual_root variable always ends in '/'. Do this with a new
"ensure_end" helper.
Signed-off-by: John Keeping <john@keeping.me.uk>
While doing this, remove declarations from header files where the
corresponding definition is declared "static" in order to avoid build
errors.
Also re-order existing headers in ui-*.c so that the file-specific
header always comes immediately after "cgit.h", helping with future
consistency.
Signed-off-by: John Keeping <john@keeping.me.uk>
Git calculates the dependency files to be included using a simply
expanded Makefile variable, so it does not include the CGit objects that
are added after that Makefile has been processed.
We therefore need to include the dependency files ourselves in order to
get the dependency calculations right. Do this.
Signed-off-by: John Keeping <john@keeping.me.uk>
This avoids needed to export every variable that might be used in
cgit.mk from the top-level Makefile.
Signed-off-by: John Keeping <john@keeping.me.uk>
This requires a small change to how we handle notes, but otherwise just
works.
Note that we can't use anything from v1.8.0 until v1.8.2.1 because some
of the symbols that we need for graph drawing were made private in
v1.8.0 and this was not reverted until v1.8.2.1.
Signed-off-by: John Keeping <john@keeping.me.uk>
It it possible to inspect blobs by specifying only the SHA-1, and CGit
provides links to do so, for example if a tag points directly at a blob.
In this case the path_items structure is never used, but creating it
still causes strlen to be run on a null pointer. Fix this.
This error was introduced by commit c1633c6 (Update git to v1.7.6.5 -
2013-03-02).
Signed-off-by: John Keeping <john@keeping.me.uk>
On some platforms (notably Solaris) /bin/sh doesn't support enough of
POSIX for gen-version.sh to run. Git's Makefile provides SHELL_PATH_SQ
to address this issue so we just have to use it.
Signed-off-by: John Keeping <john@keeping.me.uk>
If CGIT_VERSION is in CGIT_CFLAGS then a change in version (for example
because you have committed your changes) causes all of the CGit objects
to be rebuilt. Avoid this by using EXTRA_CPPFLAGS to add the version
for only those files that are affected and make them depend on VERSION.
Signed-off-by: John Keeping <john@keeping.me.uk>
We already have a global cgit_version which is set from the #define'd
CGIT_VERSION in cgit.c. Change ui-patch.c to use this so that we only
need to rebuild cgit.o when the version changes.
Signed-off-by: John Keeping <john@keeping.me.uk>
Git does quite a lot of platform-specific detection in its Makefile,
which can result in it defining preprocessor variables that are used in
its header files. If CGit does not define the same variables it can
result in different sizes of some structures in different places in the
same application.
For example, on Solaris Git uses it's "compat" regex library which has a
different sized regex_t structure than that available in the platform
regex.h. This has a knock-on effect on the size of "struct rev_info"
and leads to hard to diagnose runtime issues.
In order to avoid all of this, introduce a "cgit.mk" file that includes
Git's Makefile and make all of the existing logic apply to CGit's
objects as well. This is slightly complicated because Git's Makefile
must run in Git's directory, so all references to CGit files need to be
prefixed with "../".
In addition, OBJECTS is a simply expanded variable in Git's Makefile so
we cannot just add our objects to it. Instead we must copy the two
applicable rules into "cgit.mk". This has the advantage that we can
split CGit-specific CFLAGS from Git's CFLAGS and hence avoid rebuilding
all of Git whenever a CGit-specific value changes.
Signed-off-by: John Keeping <john@keeping.me.uk>
Acked-by: Jamie Couture <jamie.couture@gmail.com>
This ensures that the Git version pointed at by the submodule is the
same as the one that will be fetched using "make get-git".
Suggested-by: Ferry Huberts <ferry.huberts@pelagic.nl>
Signed-off-by: John Keeping <john@keeping.me.uk>
cgit_self_link() is a void function but implements each case it handles
by doing "return <another_void_function>" which is not valid C; section
6.8.6.4 of C11 says:
A return statement with an expression shall not appear in a
function whose return type is void.
Fix this by removing the return keywords and converting the final code
block into an "else".
Signed-off-by: John Keeping <john@keeping.me.uk>
The "-e" option to grep is not needed unless specifying more than one
pattern, which we don't do. Remove it to avoid restricting the tests on
platforms that do not have a grep that recognises "-e".
Signed-off-by: John Keeping <john@keeping.me.uk>
Some changes to diff options:
- no_merges has become the more general max_parents
- path restriction now uses struct pathspec
Signed-off-by: John Keeping <john@keeping.me.uk>
When a repository is empty, the ATOM feed link is written in the header,
but this involves formatting ctx->qry.head which is NULL in this case.
With glibc, vsnprintf formats "%s" with a NULL input as "(null)" but on
Solaris this results in a segmentation fault. Since we don't have a
meaningful head for the atom feed in an empty repository, it's simplest
not to write out the link element at all.
Signed-off-by: John Keeping <john@metanate.com>
John Keeping
John Keeping