Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

add psmx test to check whether shared memory can be enabled and the l… #1939

Merged

Conversation

mvandenhoek
Copy link
Contributor

…ocator can be set

@mvandenhoek mvandenhoek force-pushed the ddsc_shared_memory_test_branch branch 3 times, most recently from 74d8d92 to 2f44e5d Compare March 1, 2024 09:03
Copy link

@firuze-cagliyan firuze-cagliyan left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Review ok.

Copy link
Contributor

@dpotman dpotman left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

A few things that need a fix or can be improved, see review comments below.

cfg.psmx_instances->cfg.priority.isdefault = 1;
cfg.psmx_instances->cfg.priority.value = 1000000;
cfg.psmx_instances->next = NULL;
domain = dds_create_domain_with_rawconfig(domainId, &cfg);
Copy link
Contributor

@dpotman dpotman Mar 6, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Seems to me that using the create_participant helper function would avoid some code duplication in this test. If you'd use domain Id 5 and add an entry with a specific locator to the global psmx_locator variable, that would work I think.

bool using_psmx_iox = (strcmp(CDDS_PSMX_NAME, "iox") == 0);
ddsrt_free(CDDS_PSMX_NAME);

if ( using_psmx_iox ) {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Currently it's iox or cdds, but there could be more PSMX implementation, so I'd prefer not to assume here that we only have these 2. I think you can make the test case generic by using dds_psmx_supported_features to check for shared memory support of the plugin, and do the dds_is_shared_memory_available checks on the endpoints accordingly

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I checked it, but dds_is_shared_memory_available() internally uses dds_psmx_supported_features(). In other words, if I call dds_psmx_supported_features() first, it can't help but give the same result for dds_is_shared_memory_available(). So it seems a bit pointless to me to do it that way.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What I meant is that the else block (from line 1636) can be completely removed (and so the if on line 1376) if you parameterize the "is shared memory available". That can be done by either hard-coding that the iox implementation uses shared memory, and cdds not, or by requesting the features from the plugin. The latter will also work if we decide to add more implementations to run this test on, so I think that's the way to go. Calling dds_is_shared_memory_available() on the endpoints could be used to check if the returned value is consistent with the features the PSMX instance claims to support.

config_psmx_elm_free(cfg.psmx_instances);
}
CU_ASSERT_FATAL(domain > 0);
CU_ASSERT_FATAL(domain_has_psmx_instance_name(domain, "CycloneDDS-IOX-PSMX"));
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

You can pass an option to the iox plugin to set the SERVICE_NAME, I think it's better to use this and check for this value, so that the test does not rely on this plugin internal constant.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The SERVICE_NAME ends up in _service_name in the struct iox_psmx, which inherits from dds_psmx_t.
The iox_create_psmx() returns the struct dds_psmx, which doesn't contain the _service_name (it is in the derived class, not the base class). I couldn't find any getter functions for this. Perhaps it should be added to dds_psmx_ops_t ?

src/core/ddsc/tests/psmx.c Outdated Show resolved Hide resolved
CU_ASSERT_FATAL(dds_is_shared_memory_available(reader2));
}
{
// Write and read samples
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I wouldn't include writing and reading samples in this test, because it's about testing the is_shared_memory flag. Maybe move this to a different test case?

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I thought as 'setting the flag' only is not enough and we should also see that it works as it should be so that we can read/write samples when shared memory is enabled. Putting in another test case is also an option but it means that we should repeat some beginning steps of this test case in that one, so I think keeping these steps here is easier for us.

dds_delete(domain);
}
{
// Test case using a config without specifying a locator.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Same as above: use the helper function and avoid assumptions on plugin internals.

dds_entity_t participant = dds_create_participant(domainId, NULL, NULL);
CU_ASSERT_FATAL(participant > 0);
{
// Check that the second half of the locator is all zeros.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This assumption is based on details in the iox plugin implementation; the value assigned to the locator is the return value of the get_node_id call on the PSMX interface that returns a dds_psmx_node_identifier_t, which is 16 bytes. This happens to be an 64 bit value with trailing zero's, but could in theory be anything.

src/core/ddsc/tests/psmx.c Outdated Show resolved Hide resolved
@mvandenhoek
Copy link
Contributor Author

I think this addresses the comments, but I bumped into a problem with the cdds_psmx. When you run the same test twice in a loop (after cleaning up all entities of course, by means of dds_delete(domain);), it runs into an assert when the test code calls dds_create_topic() on the second iteration of the test.
psmx_cdds_impl.c:143: cdds_psmx_create_topic: Assertion `g_domain >= 0' failed.

The value of g_domain is -4, which stands for returncode DDS_RETCODE_PRECONDITION_NOT_MET.

The assert happens here:

static struct dds_psmx_topic * cdds_psmx_create_topic (struct dds_psmx * psmx,
const char * topic_name, const char * type_name, dds_data_type_properties_t data_type_props)
{
struct cdds_psmx *cpsmx = (struct cdds_psmx *) psmx;
if (g_domain == -1)
{
char *conf = ddsrt_expand_envvars (DDS_CONFIG, DDS_DOMAINID);
g_domain = dds_create_domain (DDS_DOMAINID, conf);
assert (g_domain >= 0);
ddsrt_free (conf);
}

The source of the returncode:

static dds_entity_t dds_domain_create_internal_xml_or_raw (dds_domain **domain_out, dds_domainid_t id, bool implicit, const struct config_source *config)
{
struct dds_domain *dom;
dds_entity_t domh = DDS_RETCODE_ERROR;
/* FIXME: should perhaps lock parent object just like everywhere */
ddsrt_mutex_lock (&dds_global.m_mutex);
retry:
if (id != DDS_DOMAIN_DEFAULT)
dom = dds_domain_find_locked (id);
else
dom = ddsrt_avl_find_min (&dds_domaintree_def, &dds_global.m_domains);
if (dom)
{
if (!implicit)
domh = DDS_RETCODE_PRECONDITION_NOT_MET;
else

It executes dom = dds_domain_find_locked (id);, then goes to domh = DDS_RETCODE_PRECONDITION_NOT_MET;.

@mvandenhoek mvandenhoek force-pushed the ddsc_shared_memory_test_branch branch from c6e507e to 0627d86 Compare April 3, 2024 11:39
@mvandenhoek
Copy link
Contributor Author

Added a dummy psmx interface, which is now used by the shared memory test.

@mvandenhoek mvandenhoek force-pushed the ddsc_shared_memory_test_branch branch 12 times, most recently from c19c331 to 00d376b Compare April 4, 2024 13:24
Copy link
Contributor

@eboasson eboasson left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thank you for this @mvandenhoek but I'm afraid I have some different ideas on how to go about testing whether the functions on the PSMX interface get invoked (and the results used) correctly by the core library. I've tried to sketch these ideas in the comments.

I won't say it has to be done the way I sketched, but on the whole I think it makes more sense the approach you chose, specifically also because it makes fewer assumptions and keeps the different tests a bit cleaner.

if(BUILD_SHARED_LIBS)
target_link_libraries(psmx_dummy PRIVATE ddsc)
endif()
install(
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We shouldn't install this, it is strictly for testing. I now see we also install the Cyclone-based one, that's also wrong.

It is somehow necessary to install it in a static build because otherwise CMake doesn't accept the reference to the psmx_dummy library when constructing libc.a, but at the same time it appears installing an OBJECT library doesn't actually install anything. In short, it seems one can get away with that.

When a shared library, installing it really copies a file, but there is also no need to install it. So I think it should be:

if(BUILD_SHARED_LIBS)
  target_link_libraries(psmx_dummy PRIVATE ddsc)
else()
  install(...)
endif()

@@ -29,6 +30,7 @@

#include "config_env.h"
#include "test_common.h"
#include "psmx_dummy_public.h"
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This include makes me think it would be better to move these new tests to a separate file. There is quite a difference between tests operating independently from the specific plugin, and tests requiring detailed knowledge of the plugin. If they are all in the same file, it takes effort to figure out the dependencies; if they are instead put in a different file, you avoid that problem.

@@ -62,17 +64,35 @@ static void fail_too_much_data (void) { fail (); }
#define TRACE_CATEGORY "discovery"
#endif

static dds_entity_t create_participant (dds_domainid_t int_dom)
/**
* @brief Create a participant object
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

So the only thing this documentation comment says that somewhat useful is NULL if not used. Clearly the function does quite a bit and so I think a documentation comment should provide a more detailed explanation of its purpose than this, or there simply shouldn't be a documentation comment at all.

I also see that the value passed in for locator is psmx_locators[...].a in all the "old" cases, and only exists for passing NULL in once and passing in some other array once. The way this wiggles its way into the function is quite ugly (for example, the -1 to handle the psmx%d bit ...)

It really is handling just three cases:

  • The old (complicated) arrangement with a plugin name taken from the environment, etc. etc.
  • The dummy-based test with no locator
  • The dummy-based test with a locator

There is a misunderstanding in what the locator string does, and as a consequence I would say distinguishing between the two latter cases is not even necessary for the purposes of this test. But that is something I'll comment upon at the new test code.

It continues to always take the plugin name from the environment (which makes some sense for when you want to control it from the outside). For the psmx_dummy-based tests, you work around that by using ddsrt_setenv to overrule the setting. That's like taking a so-so construction and making it worse.

If you look at what you actually need for the psmx_dummy plugin, it is next to nothing and so moving the dummy-based tests to a separate file with its own code for setting things up would simpler and arguably be more readable than this parameterised thing.

Copy link
Contributor Author

@mvandenhoek mvandenhoek Apr 10, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This was the least intrusive way I could think of to reuse this function. That is, without changing the interface too much, and without changing the behavior in a way that affects the other testcases. But if the dummy psmx testcase goes into another file as suggested in your earlier comment, then there is no reason to use the "ducttape approach" with ddsrt_setenv(), so the problem will solve itself.

@@ -98,6 +118,39 @@ static dds_entity_t create_participant (dds_domainid_t int_dom)
return pp;
}

static bool domain_pin(dds_entity_t domain, dds_domain** dom_out)
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't think it is worth the effort to create this function: it is used exactly twice, once for pinning the domain and checking the number of PSMX plugins, and once for getting the locator. If the second one were to operate on a dds_domain pointer it could easily be done together with the check for the number of plugins and there'd only be a single call-site.

It is also totally reasonable to require the operation to succeed. Sanity checking it is always good, but a simple:

rc = dds_entity_pin(domain, &x);
assert(rc == 0 && dds_entity_kind(x) == DDS_KIND_DOMAIN);

suffices, with the assert only being nice-to-have.

dds_entity_unpin((dds_entity*)dom);
}

static bool domain_get_psmx_locator(dds_entity_t domain, ddsi_locator_t* l_out)
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

See comment at domain_pin


//### Helper functions ###

static char* get_config_option_value (const char* conf, const char* option_name)
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'd leave this out. No need to parse the configuration, there is only a need to show that the configuration string is passed in correctly.

Everything else can easily be handled by global variables (see comments above, on accessing the statistics).


//### Dynamic library functions ###

static dummy_mockstats_t* g_mockstats;
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

So I'd simplify this by having not a global pointer to a dummy_mockstats_t, but just a global dummy_mockstats_t. (The owned one can then be deleted.)

static dummy_mockstats_t* g_mockstats;
static bool g_mockstats_owned;

static bool dummy_psmx_type_qos_supported(
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Stylistically, I think it makes more sense to define the structs with function pointers at the end, that saves a ton of function prototypes here.

dds_data_type_properties_t data_type_props
) {
(void)data_type_props;
struct dds_psmx_topic* topic = dds_alloc(sizeof(struct dds_psmx_topic));
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think it is worth considering creating two topics in the test, and mapping these to two global variables for dds_psmx_topic. That way you can extend the test application to inform the library which topic should show up in the next operation, and then verify that it does.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Am I interpreting this correctly that the test code should prepare some data in the global variable (which can be accessed by both the test code and the plugin), before calling dds_create_topic(), and then let the plugin verify from inside dummy_psmx_create_topic(), that it gets a request for the correct topic?

dds_psmx_endpoint_type_t endpoint_type
) {
(void)qos;
struct dds_psmx_endpoint* endp = dds_alloc(sizeof(struct dds_psmx_endpoint));
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Same idea as dummy_psmx_create_topic, though perhaps you'd want four of these if there are two topics.

@mvandenhoek mvandenhoek force-pushed the ddsc_shared_memory_test_branch branch from 00d376b to 99a7b81 Compare April 16, 2024 16:04
@mvandenhoek
Copy link
Contributor Author

Thank you for this @mvandenhoek but I'm afraid I have some different ideas on how to go about testing whether the functions on the PSMX interface get invoked (and the results used) correctly by the core library. I've tried to sketch these ideas in the comments.

I won't say it has to be done the way I sketched, but on the whole I think it makes more sense the approach you chose, specifically also because it makes fewer assumptions and keeps the different tests a bit cleaner.

I believe these changes address your comments.

@mvandenhoek mvandenhoek force-pushed the ddsc_shared_memory_test_branch branch 2 times, most recently from c4d55b0 to 8148bae Compare April 16, 2024 18:21

static dummy_mockstats_t g_mockstats;

static dummy_mockstats_t* dummy_mockstats_get_ptr()
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why not export this one in the same way as the two alloc functions?

dynamic_array_t endpoints;
}dummy_mockstats_t;

typedef struct dummy_psmx {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This one should be private.


static void free_strings(uint32_t len, char** strings)
{
if (len != 0 && strings != NULL) {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I generally don't care all that much about 2 vs 4 space (vs tab ...) indentation, but I do find it a bit annoying if it changes unnecessarily in one file. I would appreciate it if you can make this one use 2-space indentation as well.

{
char *configstr;
char locator_str[74];
if ( locator == NULL ) {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

These complications with locators are no longer used in any meaningful way, they should be removed. A single, fixed, configuration string suffices for the purpose of the test.

</General>\
<Discovery>\
<Tag>${CYCLONEDDS_PID}</Tag>\
<ExternalDomainId>0</ExternalDomainId>\
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This just runs a single instance of Cyclone, no need for the tricks that allow you to pretend there's a network in between when everything is in a single process.

CU_ASSERT_FATAL(reader1 > 0);
{
// Check the dummy psmx instance_name.
dds_qos_t* qos = dds_create_qos();
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why check this here, but not for the writer?

From a test structure perspective, it would also make more sense to check these in a separate test. With appropriately factored out entity creation code it takes only a few lines extra to turn it into a separate test case.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is now in a separate test.

dds_return_t rc = dds_entity_pin(reader1, &x);
CU_ASSERT_FATAL(rc == DDS_RETCODE_OK && dds_entity_kind(x) == DDS_KIND_READER);
struct dds_psmx_endpoints_set* endpt_set = &((dds_reader*)x)->m_endpoint.psmx_endpoints;
CU_ASSERT_FATAL(endpt_set->length == 1 && endpt_set->endpoints[0] == psmx_endpt_expected);
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think the only practical way to verify this in my preferred implementation is to check in the plug-in on deleting the reader. But that needs to be done anyway.

dds_entity_unpin(x);
}

create_unique_topic_name("shared_memory", topicname, sizeof(topicname));
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

All things topic2: more of the same.

// Check that shared memory is enabled when it should, and not enabled when it shouldn't.
bool psmx_enabled;
psmx_enabled = endpoint_has_psmx_enabled(writer1);
CU_ASSERT_FATAL(psmx_enabled);
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is covered already: the preceding tests are stricter than this one. (I would like those preceding ones done differently, but the net effect would still be good enough to say that this test is superfluous.)

bool psmx_enabled;
psmx_enabled = endpoint_has_psmx_enabled(writer1);
CU_ASSERT_FATAL(psmx_enabled);
CU_ASSERT_FATAL(dds_is_shared_memory_available(writer1) == supports_shared_memory_expected);
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

supports_shared_memory_expected is semantically a variable, but it is assigned only once, at declaration and its address is never taken. So it is a constant. I would prefer CU_ASSERT_FATAL(dds_is_shared_memory_available(writer1)) in such a case.

However, here, you're trying to test dds_is_shared_memory_available but with the current set-up, an implementation of dds_is_shared_memory_available that simply returns true will pass. It is therefore not a good test.

A much better way would be to:

  • Create a reader/writer that does not involve PSMX, verify that it returns false. (This would best be done in a different test.)
  • Run this test once with the plugin claiming to support shared memory and verifying it returns true. (This is the current case.)
  • Run this test once with the plugin claiming to not support shared memory and verifying it returns false. (This case is quite obviously missing.)

With the "direct" access to the dummy plugin, it is trivial to make the plugin claim one or the other.

@mvandenhoek mvandenhoek force-pushed the ddsc_shared_memory_test_branch branch from 8148bae to 22f2a7e Compare May 3, 2024 15:16
Copy link
Contributor

@eboasson eboasson left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sorry that I forgot to review this again in a reasonable time.

It is mostly in a mergeable state as far as the contents of the PR are concerned. What is still a problem is that it fails to build properly on Windows (https://dev.azure.com/eclipse-cyclonedds/fc8c2212-deca-471a-94ad-bc26de82c2b6/_apis/build/builds/6303/logs/128) because of a DLL linkage issue — that is not something caused by recent changes on master.

Another thing is that CUnit_ddsc_psmxif_config_multiple_psmx fails, presumably because the core doesn't reject the configuration as the test expects. For now, it is probably best to reject such a configuration, but I am not going to merge this PR until I have a fix in place. (If you provide a fix on this PR that would be appreciated, but otherwise I'll do it myself.)

The other good thing is that I noticed some things in the code that really need cleaning up in the core, in part because the division of labour is weird and in part because it is just written clumsily.

So it has served a purpose 🙂

@mvandenhoek
Copy link
Contributor Author

@eboasson , thanks for having a look.

It is mostly in a mergeable state as far as the contents of the PR are concerned. What is still a problem is that it fails to build properly on Windows (https://dev.azure.com/eclipse-cyclonedds/fc8c2212-deca-471a-94ad-bc26de82c2b6/_apis/build/builds/6303/logs/128) because of a DLL linkage issue — that is not something caused by recent changes on master.

Another thing is that CUnit_ddsc_psmxif_config_multiple_psmx fails, presumably because the core doesn't reject the configuration as the test expects. For now, it is probably best to reject such a configuration, but I am not going to merge this PR until I have a fix in place. (If you provide a fix on this PR that would be appreciated, but otherwise I'll do it myself.)

Yes, it fails the config multiple psmx test exactly as intended, so the test has served its purpose. I'll put these on my todo list.

…ion with multiple psmx interfaces

Signed-off-by: Michel van den Hoek <[email protected]>
@mvandenhoek mvandenhoek force-pushed the ddsc_shared_memory_test_branch branch from 947c910 to 5f904c1 Compare July 8, 2024 15:50
@mvandenhoek
Copy link
Contributor Author

It looks like all of the builds are successful now. There is one platform with one test failure unrelated to this PR.

@mvandenhoek mvandenhoek requested a review from eboasson July 9, 2024 08:19
Copy link
Contributor

@eboasson eboasson left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks @mvandenhoek! I personally would've kept the loop/sorting for loading the plugins and checked the number separately, but it is so easy to change that back when the time comes that I prefer to merging this PR now.

@eboasson eboasson merged commit f5e5d6c into eclipse-cyclonedds:master Jul 15, 2024
19 of 21 checks passed
@mvandenhoek mvandenhoek deleted the ddsc_shared_memory_test_branch branch July 15, 2024 13:30
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

4 participants