Document limits on static and dynamic linking for HPE NonStop platforms.

[rsbeckerca]

Documentation is necessary as static and dynamic linking cause SIGSEGV during atexit() processing on the platform.

Fixes: #19951

Signed-off-by: Randall S. Becker randall.becker@nexbridge.ca

Checklist

• documentation is added or updated

[t8m]

I'd replace DLL with shared library as the DLL is kind of Windows-specific acronym.

[rsbeckerca]

Will make that change imminently.

[t8m]

This should be merged to master, 3.0, and 3.1 branches.

[rsbeckerca]

This should be merged to master, 3.0, and 3.1 branches.

Do you need separate PRs for that?

[t8m]

This should be merged to master, 3.0, and 3.1 branches.

Do you need separate PRs for that?

No, I assume this will cherry-pick cleanly.

[rsbeckerca]

This should be merged to master, 3.0, and 3.1 branches.

Do you need separate PRs for that?

No, I assume this will cherry-pick cleanly.

It should. I have not mucked with that file since early 3.0.

[bernd-edlinger]

I think this description is incomplete and does not reflect the reason why
90-test-shlibload.t has been disabled on NonStop:

openssl/test/recipes/90-test_shlibload.t

Line 23 in 43a9e68

plan skip_all => "Test is disabled on NonStop" if config('target') =~ m|^nonstop|;

as far as I understand the issue, a crash will also happen if the application
does NOT link to libcrypto.so and loads a library that links to libcrypto.so

[openssl-machine]

24 hours has passed since 'approval: done' was set, but as this PR has been updated in that time the label 'approval: ready to merge' is not being automatically set. Please review the updates and set the label manually.

[t8m]

Before merging I am waiting on @rsbeckerca to respond to Bernd's comment above.

[levitte]

To be noted is that this is potentially not a NonStop problem alone...

[rsbeckerca]

I think this description is incomplete and does not reflect the reason why 90-test-shlibload.t has been disabled on NonStop:

openssl/test/recipes/90-test_shlibload.t

Line 23 in 43a9e68

plan skip_all => "Test is disabled on NonStop" if config('target') =~ m|^nonstop|;

as far as I understand the issue, a crash will also happen if the application
does NOT link to libcrypto.so and loads a library that links to libcrypto.so

The ordinals for entry points in DLLs do not line up on NonStop with the Linux linker based when this test failed in the 3.0.0 series. The requirement to have a consistent set of ordinals seems to be architecture-specific and not necessary, in my opinion. I can document this in a separate PR, if requested, but I think 90-test_shlibload has nothing to do with this specific issue.

[bernd-edlinger]

Hmm, I don't understand at all.
This test uses a stand-along executable, does dynamically load the libcrypto.so
and get the address of several exported functions by name.
calls them up. and is specifically designed to check that the atexit callback is called or not,
dependent on whether OPENSSL_init_crypto(OPENSSL_INIT_NO_ATEXIT, NULL) is called
or not.

[rsbeckerca]

Hmm, I don't understand at all. This test uses a stand-along executable, does dynamically load the libcrypto.so and get the address of several exported functions by name. calls them up. and is specifically designed to check that the atexit callback is called or not, dependent on whether OPENSSL_init_crypto(OPENSSL_INIT_NO_ATEXIT, NULL) is called or not.

The following is what happens now when I enable the test, and what happened when I reported this originally:

90-test_shlibload.t ..
# The results of this test will end up in test-runs/test_shlibload
1..10
../../util/wrap.pl ../../test/shlibloadtest -crypto_first libcrypto.so libssl.so atexit-cryptofirst.txt => 139
not ok 1 - running shlibloadtest -crypto_first atexit-cryptofirst.txt

#   Failed test 'running shlibloadtest -crypto_first atexit-cryptofirst.txt'
#   at test/recipes/90-test_shlibload.t line 34.
readline() on closed filehandle $fh at test/recipes/90-test_shlibload.t line 68.
not ok 2

#   Failed test at test/recipes/90-test_shlibload.t line 36.
../../util/wrap.pl ../../test/shlibloadtest -ssl_first libcrypto.so libssl.so atexit-sslfirst.txt => 139
not ok 3 - running shlibloadtest -ssl_first atexit-sslfirst.txt

#   Failed test 'running shlibloadtest -ssl_first atexit-sslfirst.txt'
#   at test/recipes/90-test_shlibload.t line 40.
readline() on closed filehandle $fh at test/recipes/90-test_shlibload.t line 68.
not ok 4

#   Failed test at test/recipes/90-test_shlibload.t line 42.
../../util/wrap.pl ../../test/shlibloadtest -just_crypto libcrypto.so libssl.so atexit-justcrypto.txt => 139
not ok 5 - running shlibloadtest -just_crypto atexit-justcrypto.txt

#   Failed test 'running shlibloadtest -just_crypto atexit-justcrypto.txt'
#   at test/recipes/90-test_shlibload.t line 46.
readline() on closed filehandle $fh at test/recipes/90-test_shlibload.t line 68.
not ok 6

#   Failed test at test/recipes/90-test_shlibload.t line 48.
DSO_dsobyaddr() failed
../../util/wrap.pl ../../test/shlibloadtest -dso_ref libcrypto.so libssl.so atexit-dsoref.txt => 139
not ok 7 - running shlibloadtest -dso_ref atexit-dsoref.txt

#   Failed test 'running shlibloadtest -dso_ref atexit-dsoref.txt'
#   at test/recipes/90-test_shlibload.t line 52.
readline() on closed filehandle $fh at test/recipes/90-test_shlibload.t line 68.
not ok 8

#   Failed test at test/recipes/90-test_shlibload.t line 54.
../../util/wrap.pl ../../test/shlibloadtest -no_atexit libcrypto.so libssl.so atexit-noatexit.txt => 0
ok 9 - running shlibloadtest -no_atexit atexit-noatexit.txt
readline() on closed filehandle $fh at test/recipes/90-test_shlibload.t line 68.
ok 10
# Looks like you failed 8 tests of 10.
Dubious, test returned 8 (wstat 2048, 0x800)
Failed 8/10 subtests

Test Summary Report
-------------------
90-test_shlibload.t (Wstat: 2048 Tests: 10 Failed: 8)
  Failed tests:  1-8
  Non-zero exit status: 8
Files=1, Tests=10,  1 wallclock secs ( 0.01 usr  0.00 sys +  0.58 cusr  0.01 csys =  0.61 CPU)
Result: FAIL```

[bernd-edlinger]

Okay, now this is what happens 4 times:

../../util/wrap.pl ../../test/shlibloadtest -crypto_first libcrypto.so libssl.so atexit-cryptofirst.txt => 139

139 is not the normal result code of shlibloadtest, which is 0 = ok, or 1 = "normal" error

#   at test/recipes/90-test_shlibload.t line 34.
readline() on closed filehandle $fh at test/recipes/90-test_shlibload.t line 68.

the atexit callback was not called when expected.

But 2 test cases are passed:

../../util/wrap.pl ../../test/shlibloadtest -no_atexit libcrypto.so libssl.so atexit-noatexit.txt => 0
ok 9 - running shlibloadtest -no_atexit atexit-noatexit.txt
readline() on closed filehandle $fh at test/recipes/90-test_shlibload.t line 68.
ok 10

exactly the only test case with -no_atexit, does not crash.
note the readline() on closed filehandle is a glitch, that is probably normal,
and does not make the test failed. because the test expectation is that the atexit handler
is not called and the result file does not exist as expected.

So to me this looks like exactly the same issue.

[rsbeckerca]

Okay, now this is what happens 4 times:

../../util/wrap.pl ../../test/shlibloadtest -crypto_first libcrypto.so libssl.so atexit-cryptofirst.txt => 139

139 is not the normal result code of shlibloadtest, which is 0 = ok, or 1 = "normal" error

#   at test/recipes/90-test_shlibload.t line 34.
readline() on closed filehandle $fh at test/recipes/90-test_shlibload.t line 68.

the atexit callback was not called when expected.

But 2 test cases are passed:

../../util/wrap.pl ../../test/shlibloadtest -no_atexit libcrypto.so libssl.so atexit-noatexit.txt => 0
ok 9 - running shlibloadtest -no_atexit atexit-noatexit.txt
readline() on closed filehandle $fh at test/recipes/90-test_shlibload.t line 68.
ok 10

exactly the only test case with -no_atexit, does not crash. note the readline() on closed filehandle is a glitch, that is probably normal, and does not make the test failed. because the test expectation is that the atexit handler is not called and the result file does not exist as expected.

So to me this looks like exactly the same issue.

The 139 is typically a SIGSEGV. I guess it is the same issue in that case. Do you want the two tests mentioned in the updated documentation as a result?

[bernd-edlinger]

I have no idea how to properly describe this mess :-(
Can you try out if this fixes the SIGSEGV in the shlibloadtest ?

diff --git a/test/simpledynamic.c b/test/simpledynamic.c
index 2cced8c..56fbaf2 100644
--- a/test/simpledynamic.c
+++ b/test/simpledynamic.c
@@ -22,6 +22,7 @@ int sd_load(const char *filename, SD *lib, int type)
     if (filename[strlen(filename) - 1] == ')')
         dl_flags |= RTLD_MEMBER;
 #endif
+    (void) dlopen(filename, dl_flags);
     *lib = dlopen(filename, dl_flags);
     return *lib == NULL ? 0 : 1;
 }

[t8m]

I am giving an option to put hold on this PR if somebody does it before tomorrow. Otherwise I'll merge it as it is better than nothing.

[t8m]

Merged to master, 3.1, and 3.0 branches. Thank you for your contribution.

Anybody is welcome to submit a PR that improves the text further.