curl-curl/docs/tests/TEST-SUITE.md

<!--
Copyright (C) Daniel Stenberg, <daniel@haxx.se>, et al.

SPDX-License-Identifier: curl
-->

# The curl Test Suite

# Running

See the "Requires to run" section for prerequisites.

In the root of the curl repository:

    ./configure && make && make test

To run a specific set of tests (e.g. 303 and 410):

    make test TFLAGS="303 410"

To run the tests faster, pass the -j (parallelism) flag:

    make test TFLAGS="-j10"

"make test" builds the test suite support code and invokes the 'runtests.pl'
perl script to run all the tests. The value of `TFLAGS` is passed directly
to 'runtests.pl'.

When you run tests via make, the flags `-a` and `-s` are passed, meaning to
continue running tests even after one fails, and to emit short output.

If you would like to not use those flags, you can run 'runtests.pl'
directly. You must `chdir` into the tests directory, then you can run it
like so:

    ./runtests.pl 303 410

You must have run `make test` at least once first to build the support code.

To see what flags are available for runtests.pl, and what output it emits,
run:

    man ./docs/runtests.1

After a test fails, examine the tests/log directory for stdout, stderr, and
output from the servers used in the test.

## Requires to run

- `perl` (and a Unix-style shell)
- `python` (and a Unix-style shell, for SMB and TELNET tests)
- `python-impacket` (for SMB tests)
- `diff` (when a test fails, a diff is shown)
- `stunnel` (for HTTPS and FTPS tests)
- `openssl` (the command line tool, for generating test server certificates)
- `openssh` or `SunSSH` (for SCP and SFTP tests)
- `nghttpx` (for HTTP/2 and HTTP/3 tests)

### Installation of impacket

The Python-based test servers support Python 3.

Please install python-impacket in the correct Python environment. You can
use pip or your OS' package manager to install 'impacket'.

On Debian/Ubuntu the package name is 'python3-impacket'

On FreeBSD the package name is 'py311-impacket'

On any system where pip is available: 'python3 -m pip install impacket'

You may also need to manually install the Python package 'six' as that may
be a missing requirement for impacket.

## Event-based

If curl is built with `Debug` enabled (see below), then the `runtests.pl`
script offers a `-e` option (or `--test-event`) that makes it perform
*event-based*. Such tests invokes the curl tool with `--test-event`, a
debug-only option made for this purpose.

Performing event-based means that the curl tool uses the
`curl_multi_socket_action()` API call to drive the transfer(s), instead of
the otherwise "normal" functions it would use. This allows us to test drive
the socket_action API. Transfers done this way should work exactly the same
as with the non-event based API.

To be able to use `--test-event` together with `--parallel`, curl requires
*libuv* to be present and enabled in the build: `configure --enable-libuv`

## Duplicated handles

If curl is built with `Debug` enabled (see below), then the `runtests.pl`
script offers a `--test-duphandle` option. When enabled, curl always
duplicates the easy handle and does its transfers using the new one instead
of the original. This is done entirely for testing purpose to verify that
everything works exactly the same when this is done; confirming that the
`curl_easy_duphandle()` function duplicates everything that it should.

### Port numbers used by test servers

All test servers run on "random" port numbers. All tests must be written to
use the suitable variables instead of fixed port numbers so that test cases
continue to work independently of what port numbers the test servers
actually use.

See [`FILEFORMAT`](FILEFORMAT.md) for the port number variables.

### Test servers

The test suite runs stand-alone servers on random ports to which it makes
requests. For SSL tests, it runs stunnel to handle encryption to the regular
servers. For SSH, it runs a standard OpenSSH server.

The listen port numbers for the test servers are picked randomly to allow
users to run multiple test cases concurrently and to not collide with other
existing services that might listen to ports on the machine.

The HTTP server supports listening on a Unix domain socket, the default
location is 'http.sock'.

For HTTP/2 and HTTP/3 testing an installed `nghttpx` is used. HTTP/3 tests
check if nghttpx supports the protocol. To override the nghttpx used, set
the environment variable `NGHTTPX`. The default can also be changed by
specifying `--with-test-nghttpx=<path>` as argument to `configure`.

### DNS server

There is a test DNS server to allow tests to resolve hostnames to verify
those code paths. This server is started like all the other servers within
the `<servers>` section.

To make a curl build actually use the test DNS server requires a debug
build. When such a test runs, the environment variable `CURL_DNS_SERVER` is
set to identify the IP address and port number of the DNS server to use.

- curl built to use c-ares for resolving automatically asks that server for
  host information

- curl built to use `getaddrinfo()` for resolving *and* is built with c-ares
  1.26.0 or later, gets a special work-around. In such builds, when the
  environment variable is set, curl instead invokes a getaddrinfo wrapper
  that emulates the function and acknowledges the DNS server environment
  variable. This way, the getaddrinfo-using code paths in curl are verified,
  and yet the custom responses from the test DNS server are used.

curl that is built to support a custom DNS server in a test gets the
`override-dns` feature set.

When curl ask for HTTPS-RR, c-ares is always used and in debug builds such
asks respects the dns server environment variable as well.

The test DNS server only has a few limited responses. When asked for

- type `A` response, it returns the address `127.0.0.1` three times
- type `AAAA` response, it returns the address `::1` three times
- other types, it returns a blank response without answers

### Shell startup scripts

Tests which use the ssh test server, SCP/SFTP tests, might be badly
influenced by the output of system wide or user specific shell startup
scripts, .bashrc, .profile, /etc/csh.cshrc, .login, /etc/bashrc, etc. which
output text messages or escape sequences on user login. When these shell
startup messages or escape sequences are output they might corrupt the
expected stream of data which flows to the sftp-server or from the ssh
client which can result in bad test behavior or even prevent the test server
from running.

If the test suite ssh or sftp server fails to start up and logs the message
'Received message too long' then you are certainly suffering the unwanted
output of a shell startup script. Locate, cleanup or adjust the shell
script.

### Memory test

The test script checks that all allocated memory is freed properly IF curl
has been built with the `DEBUGBUILD` define set. The script automatically
detects if that is the case, and it uses the `memanalyze.pl` script to
analyze the memory debugging output.

Also, if you run tests on a machine where valgrind is found, the script uses
valgrind to run the test with (unless you use `-n`) to further verify
correctness.

The `runtests.pl` `-t` option enables torture testing mode. It runs each
test many times and makes each different memory allocation fail on each
successive run. This tests the out of memory error handling code to ensure
that memory leaks do not occur even in those situations. It can help to
compile curl with `CPPFLAGS=-DMEMDEBUG_LOG_SYNC` when using this option, to
ensure that the memory log file is properly written even if curl crashes.

### Debug

If a test case fails, you can conveniently get the script to invoke the
debugger (gdb) for you with the server running and the same command line
parameters that failed. Invoke `runtests.pl <test number> -g` and then
type 'run' in the debugger to perform the command through the debugger.

### Logs

All logs are generated in the log/ subdirectory (it is emptied first in the
runtests.pl script). They remain in there after a test run.

### Log Verbosity

A curl build with `--enable-debug` offers more verbose output in the logs.
This applies not only for test cases, but also when running it standalone
with `curl -v`. While a curl debug built is
***not suitable for production***, it is often helpful in tracking down
problems.

Sometimes, one needs detailed logging of operations, but does not want
to drown in output. The newly introduced *connection filters* allows one to
dynamically increase log verbosity for a particular *filter type*. Example:

    CURL_DEBUG=ssl curl -v https://curl.se/

makes the `ssl` connection filter log more details. One may do that for
every filter type and also use a combination of names, separated by `,` or
space.

    CURL_DEBUG=ssl,http/2 curl -v https://curl.se/

The order of filter type names is not relevant. Names used here are
case insensitive. Note that these names are implementation internals and
subject to change.

Some, likely stable names are `tcp`, `ssl`, `http/2`. For a current list,
one may search the sources for `struct Curl_cftype` definitions and find
the names there. Also, some filters are only available with certain build
options, of course.

### Test input files

All test cases are put in the `data/` subdirectory. Each test is stored in
the file named according to the test number.

See [`FILEFORMAT`](FILEFORMAT.md) for a description of the test case file
format.

### Code coverage

gcc provides a tool that can determine the code coverage figures for the
test suite. To use it, configure curl with `CFLAGS='-fprofile-arcs
-ftest-coverage -g -O0'`. Make sure you run the normal and torture tests to
get more full coverage, i.e. do:

    make test
    make test-torture

The graphical tool `ggcov` can be used to browse the source and create
coverage reports on \*nix hosts:

    ggcov -r lib src

The text mode tool `gcov` may also be used, but it does not handle object
files in more than one directory correctly.

### Remote testing

The runtests.pl script provides some hooks to allow curl to be tested on a
machine where perl can not be run. The test framework in this case runs on
a workstation where perl is available, while curl itself is run on a remote
system using ssh or some other remote execution method. See the comments at
the beginning of runtests.pl for details.

## Test case numbering

Test cases used to be numbered by category ranges, but the ranges filled
up. Subsets of tests can now be selected by passing keywords to the
runtests.pl script via the make `TFLAGS` variable.

New tests are added by finding a free number in `tests/data/Makefile.am`.

## Write tests

Here's a quick description on writing test cases. We have three kinds of
tests: the ones that test the curl tool, the ones that build small
applications and test libcurl directly and the unit tests that test individual
(possibly internal) functions.

### test data

Each test has a master file that controls all the test data. What to read,
what the protocol exchange should look like, what exit code to expect and
what command line arguments to use etc.

These files are `tests/data/test[num]` where `[num]` is a unique identifier
described above, and the XML-like file format of them is described in the
separate [`FILEFORMAT`](FILEFORMAT.md) document.

### curl tests

A test case that runs the curl tool and verifies that it gets the correct
data, it sends the correct data, it uses the correct protocol primitives
etc.

### libcurl tests

The libcurl tests are identical to the curl ones, except that they use a
specific and dedicated custom-built program to run instead of "curl". This
tool is built from source code placed in `tests/libtest` and if you want to
make a new libcurl test that is where you add your code.

### unit tests

Unit tests are placed in `tests/unit`. There is a tests/unit/README
describing the specific set of checks and macros that may be used when
writing tests that verify behaviors of specific individual functions.

The unit tests depend on curl being built with debug enabled.

### test bundles

Individual tests are bundled into single executables, one for libtests, one
for unit tests and one for servers. The executables' first argument is
the name of libtest, unit test or server respectively.
In these executables, the build process automatically renames the entry point
to a unique symbol. `test` becomes `test_<tool>`, e.g. `test_lib1598` or
`test_unit1305`. For servers `main` becomes `main_sws` for the `sws` server,
and so on. Other common symbols may also be suffixed the same way.