maria-developers team mailing list archive

[Teodor Mircea Ionita <teodor@xxxxxxxxxxx>]

Hi,

While trying to build on macOS 10.13.4 with clang 9.1.0 I originally came across this error:
--
.../storage/innobase/include/sync0policy.h:79:4: error: cannot initialize a parameter of type 'int64' (aka 'long long') with
     an rvalue of type 'os_thread_id_t' (aka '_opaque_pthread_t *')
                       my_atomic_storelint(&m_thread_id, os_thread_get_curr_id());
---
This problem was originally reported in MDEV-15778. Since pthread_t on macOS is a pointer to a struct named  _opaque_pthread_t, the "obvious" solution for this was to explicitly cast to ULINT which in turn is defined as size_t, fine so far. Solution was tested in PR#691.

However, upon testing my patch with buildbot, on slave kvm-fulltest2, which is 64bit qemu vm with Ubuntu precise-i386 (32bit) running and gcc compiling in -march=i686. The above "fix" now yields:
--
/home/buildbot/buildbot/build/mariadb-10.3.6/storage/innobase/include/sync0policy.h:79:4: error: invalid cast from type ‘os_thread_id_t {aka long unsigned int}’ to type ‘ulint {aka unsigned int}’
---

Now reconsidering more carefully, this looks more serious then initially thought, posix pthread_t(3) is defined as:
--
POSIX.1 allows an implementation wide freedom in choosing the type used to represent a thread ID; for example, representation using either an arithmetic type or a structure is permitted. Therefore, variables of type pthread_t can't portably be compared using the C equality operator (==); use pthread_equal(3)instead.
---

So in this particular platform spec, where pthread_t on Linux is defined as unsigned long int integral type and if LP64 model is used, this implies pthread_t always being 64bit in either both 32 and 64 arch and trying to convert to size_t on a 32bit arch will fail.

Now given that the standard particularly states pthread_t as being opaque and no assumptions should be made in regards to the actual type, not even an initializer type (which ULINT_UNDEFINED defines as -1), trying to cast m_thread_id to any other type ( (ULINT) ) either implicitly or explicit is not portable. Is this a bug that needs fixing or just impossible to handle pthread id atomically in any other fashion? How about a test case I can run to confirm/infirm erratic behavior?

Since I don't have access to the particular bb slave instance to check with sizeof() maybe somebody else can help me here to confirm this? Relevant links bellow:

https://jira.mariadb.org/browse/MDEV-15778
https://buildbot.askmonty.org/buildbot/builders/kvm-fulltest2/builds/12317/steps/compile/logs/stdio
https://linux.die.net/man/3/pthread_self
https://github.com/MariaDB/server/pull/691

Cheers,
Teodor