doc: Explain briefly how to write new tests

Add a second on writing tests, covering when to use Python and C, where
to put the tests, etc. Add a link to the existing Python test
documentation.

Signed-off-by: Simon Glass <sjg@chromium.org>

doc/develop/index.rst

@@ -33,4 +33,5 @@ Testing
    coccinelle
    testing
    py_testing
+   tests_writing
    tests_sandbox

doc/develop/py_testing.rst

@@ -13,7 +13,8 @@ results. Advantages of this approach are:
   U-Boot; there can be no disconnect.
 - There is no need to write or embed test-related code into U-Boot itself.
   It is asserted that writing test-related code in Python is simpler and more
-  flexible than writing it all in C.
+  flexible than writing it all in C. But see :doc:`tests_writing` for caveats
+  and more discussion / analysis.
 - It is reasonably simple to interact with U-Boot in this way.
 
 Requirements

doc/develop/testing.rst

@@ -117,6 +117,8 @@ or is covered sparingly. So here are some suggestions:
   is much easier to add onto a test - writing a new large test can seem
   daunting to most contributors.
 
+See doc:`tests_writing` for how to write tests.
+
 
 Future work
 -----------

doc/develop/tests_sandbox.rst

@@ -200,3 +200,10 @@ linker_list::
    000000000001f240 D _u_boot_list_2_dm_test_2_dm_test_of_plat_parent
    000000000001f260 D _u_boot_list_2_dm_test_2_dm_test_of_plat_phandle
    000000000001f280 D _u_boot_list_2_dm_test_2_dm_test_of_plat_props
+
+
+Writing tests
+-------------
+
+See :doc:`tests_writing` for how to write new tests.
+

doc/develop/tests_writing.rst

@@ -0,0 +1,346 @@
+.. SPDX-License-Identifier: GPL-2.0+
+.. Copyright 2021 Google LLC
+.. sectionauthor:: Simon Glass <sjg@chromium.org>
+
+Writing Tests
+=============
+
+This describes how to write tests in U-Boot and describes the possible options.
+
+Test types
+----------
+
+There are two basic types of test in U-Boot:
+
+  - Python tests, in test/py/tests
+  - C tests, in test/ and its subdirectories
+
+(there are also UEFI tests in lib/efi_selftest/ not considered here.)
+
+Python tests talk to U-Boot via the command line. They support both sandbox and
+real hardware. They typically do not require building test code into U-Boot
+itself. They are fairly slow to run, due to the command-line interface and there
+being two separate processes. Python tests are fairly easy to write. They can
+be a little tricky to debug sometimes due to the voluminous output of pytest.
+
+C tests are written directly in U-Boot. While they can be used on boards, they
+are more commonly used with sandbox, as they obviously add to U-Boot code size.
+C tests are easy to write so long as the required facilities exist. Where they
+do not it can involve refactoring or adding new features to sandbox. They are
+fast to run and easy to debug.
+
+Regardless of which test type is used, all tests are collected and run by the
+pytest framework, so there is typically no need to run them separately. This
+means that C tests can be used when it makes sense, and Python tests when it
+doesn't.
+
+
+This table shows how to decide whether to write a C or Python test:
+
+=====================  ===========================  =============================
+Attribute              C test                       Python test
+=====================  ===========================  =============================
+Fast to run?           Yes                          No (two separate processes)
+Easy to write?         Yes, if required test        Yes
+                       features exist in sandbox
+                       or the target system
+Needs code in U-Boot?  Yes                          No, provided the test can be
+                                                    executed and the result
+                                                    determined using the command
+                                                    line
+Easy to debug?         Yes                          No, since access to the U-Boot
+                                                    state is not available and the
+                                                    amount of output can
+                                                    sometimes require a bit of
+                                                    digging
+Can use gdb?           Yes, directly                Yes, with --gdbserver
+Can run on boards?     Some can, but only if        Some
+                       compiled in and not
+                       dependent on sandboxau
+=====================  ===========================  =============================
+
+
+Python or C
+-----------
+
+Typically in U-Boot we encourage C test using sandbox for all features. This
+allows fast testing, easy development and allows contributors to make changes
+without needing dozens of boards to test with.
+
+When a test requires setup or interaction with the running host (such as to
+generate images and then running U-Boot to check that they can be loaded), or
+cannot be run on sandbox, Python tests should be used. These should typically
+NOT rely on running with sandbox, but instead should function correctly on any
+board supported by U-Boot.
+
+
+How slow are Python tests?
+--------------------------
+
+Under the hood, when running on sandbox, Python tests work by starting a sandbox
+test and connecting to it via a pipe. Each interaction with the U-Boot process
+requires at least a context switch to handle the pipe interaction. The test
+sends a command to U-Boot, which then reacts and shows some output, then the
+test sees that and continues. Of course on real hardware, communications delays
+(e.g. with a serial console) make this slower.
+
+For comparison, consider a test that checks the 'md' (memory dump). All times
+below are approximate, as measured on an AMD 2950X system. Here is is the test
+in Python::
+
+   @pytest.mark.buildconfigspec('cmd_memory')
+   def test_md(u_boot_console):
+       """Test that md reads memory as expected, and that memory can be modified
+       using the mw command."""
+
+       ram_base = u_boot_utils.find_ram_base(u_boot_console)
+       addr = '%08x' % ram_base
+       val = 'a5f09876'
+       expected_response = addr + ': ' + val
+       u_boot_console.run_command('mw ' + addr + ' 0 10')
+       response = u_boot_console.run_command('md ' + addr + ' 10')
+       assert(not (expected_response in response))
+       u_boot_console.run_command('mw ' + addr + ' ' + val)
+       response = u_boot_console.run_command('md ' + addr + ' 10')
+       assert(expected_response in response)
+
+This runs a few commands and checks the output. Note that it runs a command,
+waits for the response and then checks it agains what is expected. If run by
+itself it takes around 800ms, including test collection. For 1000 runs it takes
+19 seconds, or 19ms per run. Of course 1000 runs it not that useful since we
+only want to run it once.
+
+There is no exactly equivalent C test, but here is a similar one that tests 'ms'
+(memory search)::
+
+   /* Test 'ms' command with bytes */
+   static int mem_test_ms_b(struct unit_test_state *uts)
+   {
+      u8 *buf;
+
+      buf = map_sysmem(0, BUF_SIZE + 1);
+      memset(buf, '\0', BUF_SIZE);
+      buf[0x0] = 0x12;
+      buf[0x31] = 0x12;
+      buf[0xff] = 0x12;
+      buf[0x100] = 0x12;
+      ut_assertok(console_record_reset_enable());
+      run_command("ms.b 1 ff 12", 0);
+      ut_assert_nextline("00000030: 00 12 00 00 00 00 00 00 00 00 00 00 00 00 00 00    ................");
+      ut_assert_nextline("--");
+      ut_assert_nextline("000000f0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 12    ................");
+      ut_assert_nextline("2 matches");
+      ut_assert_console_end();
+
+      ut_asserteq(2, env_get_hex("memmatches", 0));
+      ut_asserteq(0xff, env_get_hex("memaddr", 0));
+      ut_asserteq(0xfe, env_get_hex("mempos", 0));
+
+      unmap_sysmem(buf);
+
+      return 0;
+   }
+   MEM_TEST(mem_test_ms_b, UT_TESTF_CONSOLE_REC);
+
+This runs the command directly in U-Boot, then checks the console output, also
+directly in U-Boot. If run by itself this takes 100ms. For 1000 runs it takes
+660ms, or 0.66ms per run.
+
+So overall running a C test is perhaps 8 times faster individually and the
+interactions are perhaps 25 times faster.
+
+It should also be noted that the C test is fairly easy to debug. You can set a
+breakpoint on do_mem_search(), which is what implements the 'ms' command,
+single step to see what might be wrong, etc. That is also possible with the
+pytest, but requires two terminals and --gdbserver.
+
+
+Why does speed matter?
+----------------------
+
+Many development activities rely on running tests:
+
+  - 'git bisect run make qcheck' can be used to find a failing commit
+  - test-driven development relies on quick iteration of build/test
+  - U-Boot's continuous integration (CI) systems make use of tests. Running
+      all sandbox tests typically takes 90 seconds and running each qemu test
+      takes about 30 seconds. This is currently dwarfed by the time taken to
+      build all boards
+
+As U-Boot continues to grow its feature set, fast and reliable tests are a
+critical factor factor in developer productivity and happiness.
+
+
+Writing C tests
+---------------
+
+C tests are arranged into suites which are typically executed by the 'ut'
+command. Each suite is in its own file. This section describes how to accomplish
+some common test tasks.
+
+(there are also UEFI C tests in lib/efi_selftest/ not considered here.)
+
+Add a new driver model test
+~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Use this when adding a test for a new or existing uclass, adding new operations
+or features to a uclass, adding new ofnode or dev_read_() functions, or anything
+else related to driver model.
+
+Find a suitable place for your test, perhaps near other test functions in
+existing code, or in a new file. Each uclass should have its own test file.
+
+Declare the test with::
+
+   /* Test that ... */
+   static int dm_test_uclassname_what(struct unit_test_state *uts)
+   {
+      /* test code here */
+
+      return 0;
+   }
+   DM_TEST(dm_test_uclassname_what, UT_TESTF_SCAN_FDT);
+
+Replace 'uclassname' with the name of your uclass, if applicable. Replace 'what'
+with what you are testing.
+
+The flags for DM_TEST() are defined in test/test.h and you typically want
+UT_TESTF_SCAN_FDT so that the devicetree is scanned and all devices are bound
+and ready for use. The DM_TEST macro adds UT_TESTF_DM automatically so that
+the test runner knows it is a driver model test.
+
+Driver model tests are special in that the entire driver model state is
+recreated anew for each test. This ensures that if a previous test deletes a
+device, for example, it does not affect subsequent tests. Driver model tests
+also run both with livetree and flattree, to ensure that both devicetree
+implementations work as expected.
+
+Example commit: c48cb7ebfb4 ("sandbox: add ADC unit tests") [1]
+
+[1] https://gitlab.denx.de/u-boot/u-boot/-/commit/c48cb7ebfb4
+
+
+Add a C test to an existing suite
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Use this when you are adding to or modifying an existing feature outside driver
+model. An example is bloblist.
+
+Add a new function in the same file as the rest of the suite and register it
+with the suite. For example, to add a new mem_search test::
+
+   /* Test 'ms' command with 32-bit values */
+   static int mem_test_ms_new_thing(struct unit_test_state *uts)
+   {
+         /* test code here*/
+
+         return 0;
+   }
+   MEM_TEST(mem_test_ms_new_thing, UT_TESTF_CONSOLE_REC);
+
+Note that the MEM_TEST() macros is defined at the top of the file.
+
+Example commit: 9fe064646d2 ("bloblist: Support relocating to a larger space") [1]
+
+[1] https://gitlab.denx.de/u-boot/u-boot/-/commit/9fe064646d2
+
+
+Add a new test suite
+~~~~~~~~~~~~~~~~~~~~
+
+Each suite should focus on one feature or subsystem, so if you are writing a
+new one of those, you should add a new suite.
+
+Create a new file in test/ or a subdirectory and define a macro to register the
+suite. For example::
+
+   #include <common.h>
+   #include <console.h>
+   #include <mapmem.h>
+   #include <dm/test.h>
+   #include <test/ut.h>
+
+   /* Declare a new wibble test */
+   #define WIBBLE_TEST(_name, _flags)   UNIT_TEST(_name, _flags, wibble_test)
+
+   /* Tetss go here */
+
+   /* At the bottom of the file: */
+
+   int do_ut_wibble(struct cmd_tbl *cmdtp, int flag, int argc, char *const argv[])
+   {
+     struct unit_test *tests = UNIT_TEST_SUITE_START(wibble_test);
+     const int n_ents = UNIT_TEST_SUITE_COUNT(wibble_test);
+
+     return cmd_ut_category("cmd_wibble", "wibble_test_", tests, n_ents, argc, argv);
+   }
+
+Then add new tests to it as above.
+
+Register this new suite in test/cmd_ut.c by adding to cmd_ut_sub[]::
+
+  /* Within cmd_ut_sub[]... */
+
+  U_BOOT_CMD_MKENT(wibble, CONFIG_SYS_MAXARGS, 1, do_ut_wibble, "", ""),
+
+and adding new help to ut_help_text[]::
+
+  "ut wibble - Test the wibble feature\n"
+
+If your feature is conditional on a particular Kconfig, then you can use #ifdef
+to control that.
+
+Finally, add the test to the build by adding to the Makefile in the same
+directory::
+
+  obj-$(CONFIG_$(SPL_)CMDLINE) += wibble.o
+
+Note that CMDLINE is never enabled in SPL, so this test will only be present in
+U-Boot proper. See below for how to do SPL tests.
+
+As before, you can add an extra Kconfig check if needed::
+
+  ifneq ($(CONFIG_$(SPL_)WIBBLE),)
+  obj-$(CONFIG_$(SPL_)CMDLINE) += wibble.o
+  endif
+
+
+Example commit: 919e7a8fb64 ("test: Add a simple test for bloblist") [1]
+
+[1] https://gitlab.denx.de/u-boot/u-boot/-/commit/919e7a8fb64
+
+
+Making the test run from pytest
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+All C tests must run from pytest. Typically this is automatic, since pytest
+scans the U-Boot executable for available tests to run. So long as you have a
+'ut' subcommand for your test suite, it will run. The same applies for driver
+model tests since they use the 'ut dm' subcommand.
+
+See test/py/tests/test_ut.py for how unit tests are run.
+
+
+Add a C test for SPL
+~~~~~~~~~~~~~~~~~~~~
+
+Note: C tests are only available for sandbox_spl at present. There is currently
+no mechanism in other boards to existing SPL tests even if they are built into
+the image.
+
+SPL tests cannot be run from the 'ut' command since there are no commands
+available in SPL. Instead, sandbox (only) calls ut_run_list() on start-up, when
+the -u flag is given. This runs the available unit tests, no matter what suite
+they are in.
+
+To create a new SPL test, follow the same rules as above, either adding to an
+existing suite or creating a new one.
+
+An example SPL test is spl_test_load().
+
+
+Writing Python tests
+--------------------
+
+See :doc:`py_testing` for brief notes how to write Python tests. You
+should be able to use the existing tests in test/py/tests as examples.