PETSc module initialization occurs before MPI can be initialized, causing segfault

[sloisel]

Hi guys, thanks for this library. I've been using it myself but I encountered a bug. Not sure if this is PETSc.jl's fault but I had Claude draft me an issue, I'm sure you'll be able to determine if this is really PETSc.jl's fault or if I am to blame!

PETSc module initialization occurs before MPI can be initialized, causing segfault

Description

When using PETSc.jl in an MPI environment, the module's __init__() function runs during using PETSc, which occurs before user code can call MPI.Init(). This causes segfaults because PETSc's initialization tries to access MPI functionality before MPI is properly initialized.

This is particularly problematic when PETSc.jl is loaded as a dependency of another package, as users have no control over when using PETSc executes.

Environment

• Julia version: 1.11.7
• PETSc.jl version: 0.3.1
• MPI.jl version: 0.20.23
• OS: macOS (Darwin 25.1.0)

Minimal Reproduction

File: test_petsc_segfault.jl

using MPI

# This is the problematic sequence:
# 1. Load PETSc BEFORE calling MPI.Init()
using PETSc  # <-- PETSc.__init__() runs here, tries to access MPI

# 2. Initialize MPI (too late!)
if !MPI.Initialized()
    MPI.Init()
end

# 3. Initialize PETSc
if !PETSc.initialized(PETSc.petsclibs[1])
    PETSc.initialize()
end

println("If we got here, it worked!")

Run with:

mpiexec -n 2 julia test_petsc_segfault.jl

Expected result: Program runs without error

Actual result: Segmentation fault

Root Cause

The execution sequence is:

1. mpiexec launches Julia → MPI environment is set up by mpiexec
2. User code: using MPI
3. User code: using PETSc
   • PETSc.jl's __init__() runs here, potentially accessing MPI before it's initialized
4. User code: MPI.Init() is called (too late!)
5. User code: PETSc.initialize() is called

Current Workaround

The only workaround is to ensure MPI.Init() is called before using PETSc:

File: test_petsc_workaround.jl

using MPI

# WORKAROUND: Initialize MPI BEFORE loading PETSc
if !MPI.Initialized()
    MPI.Init()
end

# Now safe to load PETSc
using PETSc

# Initialize PETSc
if !PETSc.initialized(PETSc.petsclibs[1])
    PETSc.initialize()
end

println("Success with workaround!")

Impact

This issue makes it nearly impossible to use PETSc.jl as a dependency in other packages, because:

1. When a user does using MyPackage (which depends on PETSc.jl), the using PETSc happens during module loading
2. This occurs before the user's test file can call MPI.Init()
3. Result: immediate segfault

Suggested Fix

PETSc.jl's __init__() function should avoid accessing MPI functionality, or should check if MPI is initialized first. The actual PETSc initialization should be deferred to an explicit call to PETSc.initialize(), which users already call after MPI.Init().

Alternatively, document that PETSc.jl should only be loaded after MPI.Init() has been called, though this would require significant restructuring of code that depends on PETSc.jl.

[sloisel]

Actually this is claude being an idiot, nevermind.