Struct PboClientParts 

pub struct PboClientParts {
    pub state: EfClientState,
    pub map_fn: EfPboMapFn,
    pub unmap_fn: EfPboUnmapFn,
    pub buffer_id: u32,
    pub size: usize,
    pub shape: Vec<usize>,
    /* private fields */
}

The C-ABI parts ef_tensor_wrap_pbo takes, from a PboTensor at the origin of its buffer.

“At the origin”, not “whole-buffer”: what PboTensor::into_client_parts guards is the byte offset, not the extent. view(0, &[16]) on a 32-byte buffer passes and yields shape = [16], size = 32 — a correct, representable tensor, because size is the GL allocation’s full byte count and is documented to exceed the shape’s product (that is exactly how a stride-padded PBO is carried). A non-zero view_offset is the only thing this struct cannot express, and that is what is refused.

state.ctx borrows the channel this value still owns; the callee is expected to retain it (which client_state_pbo_ops does) before this struct drops. Dropping it without retaining releases the client’s own last reference and deletes the GL buffer — which is why the Arc is a field here rather than something the caller has to remember to hold.

Deliberately neither Send nor Sync: EfClientState holds a raw pointer and edgefirst-tensor-abi adds no unsafe impl. That is right for a value whose only purpose is to be handed straight to a C entry point on the calling thread. If this ever needs to cross a thread, that is a design question to raise, not something to unblock with an unsafe impl.

Fields

state: EfClientState

The callback channel, borrowing this value’s own reference.

map_fn: EfPboMapFn

unmap_fn: EfPboUnmapFn

buffer_id: u32

size: usize

The GL allocation’s full byte count, which may exceed the shape’s product (a 64-byte-aligned row stride).

shape: Vec<usize>