Re: RFC: Proposal: Add NUMA Memory Policy Tunable to glibc malloc implementation

[罗春生]

Dear Florian,

>I don't see why you need LD_PRELOAD for this.  The manual page says that
>set_mempolicy is preserved across fork and execve.  For such process
>attributes, we do not provide tunables.

Thank you for your insightful feedback.

My reference to LD_PRELOAD was strictly regarding workaround scenarios where developers might 
try to intercept early allocations through malloc hooks before main() executes. For example:
// preload_hack.c  
__attribute__((constructor)) void malloc_hook(int size) {  
    // Attempt to set policy before libc constructors without allocating additional memory  
    set_mempolicy(MPOL_INTERLEAVE, numa_nodes_mask);  
    return origin_malloc(size);  
}
This approach has known limitations: reentrancy risks during dynamic linking and potential performence issues.
Our proposal aims to provide a standardized alternative to such brittle solutions.

I completely agree with your point about the inheritable nature of set_mempolicy()'s process attributes - 
the memory policy persists across execve() and is inherited by child processes (as documented in the manual). 
This tunable implementation would affect all memory allocation policies of a process and its children, not just ptmalloc's behavior.

The original proposal addressed a specific timing challenge: critical allocations occurring during the initialization sequence:
ld.so initialization → early malloc calls → first user call to set_mempolicy()

After careful consideration, we realized that might not have a real problem - we may have over-engineered a solution for a non-critical situation.

Thank you for your expert guidance.

Best regards,
Chunsheng Luo