Segfault in hsa_init on non-HSA platforms

[r-potter]

Prior to 1.0.3, calling hsa_init() on a non-HSA platform (such as an Intel CPU) was well behaved. I believe the behaviour was that hsa_init() succeeded, but that the runtime would report no agents or regions to the hsa_iterate_*.

With the 1.0.3 release, hsa_init segfaults. This is due to this line:

// Setup system region allocator.
if (reinterpret_castamd::MemoryRegion*(
core::MemoryRegion::Convert(system_region_))->fine_grain()) {

On non-HSA platforms, system_region_.handle == 0, so the call to fine_grain() dereferences a null pointer.

Having hsa_init() fail cleanly, or restoring the previous behaviour, where hsa_init() succeeds, but no agents and regions are reported would be preferable to a segfault.

It appears that simply adding a check for a null system_region_ restores this behaviour.

runtime.patch.txt

[tschwinge]

Confirming the problem (master branch: commit 576359e "HSA 1.0.3 Source"):

[...]
Program received signal SIGSEGV, Segmentation fault.
0x00007ffff564e579 in core::Runtime::Load() () from [...]/libhsa-runtime64.so.1
(gdb) bt
#0  0x00007ffff564e579 in core::Runtime::Load() () from [...]/libhsa-runtime64.so.1
#1  0x00007ffff564d0ee in core::Runtime::Acquire() () from [...]/libhsa-runtime64.so.1
#2  0x00007ffff563910a in HSA::hsa_init() () from [...]/libhsa-runtime64.so.1
#3  0x00007ffff565e074 in hsa_init () from [...]/libhsa-runtime64.so.1
[...]

..., and that the patch posted by @r-potter, resolves this.

[jedwards-AMD]

This will be fixed in the next release.