Crash on PkgEval

[maleadt]

ArchGDAL.jl fails tests on PkgEval, specifically, it segfaults:

[355] signal (11.1): Segmentation fault
in expression starting at /home/pkgeval/.julia/packages/ArchGDAL/AZONk/test/test_tables.jl:5
OGR_FD_GetFieldIndex at /home/pkgeval/.julia/artifacts/131cab42093e41cb0d9c0382948d9aa9998b33cb/lib/libgdal.so (unknown line)
#createlayer#236 at /home/pkgeval/.julia/packages/ArchGDAL/AZONk/src/context.jl:267
createlayer at /home/pkgeval/.julia/packages/ArchGDAL/AZONk/src/context.jl:264 [inlined]
#14 at /home/pkgeval/.julia/packages/ArchGDAL/AZONk/test/test_tables.jl:268 [inlined]
#create#226 at /home/pkgeval/.julia/packages/ArchGDAL/AZONk/src/context.jl:267

This doesn't seem like a Julia-related crash, as it crashes during libgdal.so. As this crash has been on the top of PkgEval reports for a while, it would be nice to see this fixed. You can easily reproduce the PkgEval environment (on Linux) by installing PkgEval from https://github.com/JuliaCI/PkgEval.jl and doing:

julia> using PkgEval

julia> config = Configuration(julia="nightly")
PkgEval configuration 'unnamed' (
  - julia: nightly
...
)

julia> PkgEval.sandboxed_julia(config)

# this spawns a sandbox where you can install and test ArchGDAL

(@v1.10) pkg> add ArchGDAL

(@v1.10) pkg> test ArchGDAL

[355] signal (11.1): Segmentation fault

[maleadt]

I see there's a PR already, great! I've blacklisted testing of ArchGDAL for now, just to reduce the noise on PkgEval reports; so could you ping (or create a PR to remove the blacklist) when the fix is part of a tagged release? Thanks!

[visr]

Will do! Turns out it what is really needed is lots of f(obj.ptr) to GC.@preserve obj f(obj.ptr) all over the package, which I added now in #348.

[maleadt]

Ah yes, the GC has become quite a bit more aggressive lately so it's likely that any missing lifetime annotations will cause corruption now.

[maleadt]

Note that generally though, you don't need these, only if you're explicitly deriving a raw resource (e.g. a pointer) from a tracked object. Normally, that conversion should happen by Base.unsafe_convert as called by ccall, in which case the Julia compiler will track lifetimes accordingly.

[visr]

Interesting. ArchGDAL.jl defines structs (say Dataset) that wrap pointers, to create a convenient typed interface on top of GDAL.jl, which has the ccalls. So at the point that ArchGDAL calls into GDAL it always has to take out the pointer. Would it be ok to just define Base.unsafe_convert(::Type{Ptr{Void}, x::Dataset) = x.ptr, such that ccall knows` how to take out the pointer whilst tracking lifetimes? That would be an easier solution.

[maleadt]

If the contained pointers' lifetime is tracked by the Dataset object, then yes, having unsafe_convert return the pointer would be sufficient.

[visr]

Should be all fixed now on ArchGDAL v0.9.4, I put up https://github.com/JuliaCI/PkgEval.jl/pull/203 to remove it from the skip list.