Can pw_kernel Avoid IPC Copy Overhead for Large-Buffer Firmware Workflows?

[rusty1968]

IPC Copy Overhead in Large-Buffer Firmware Workflows

Firmware measurement and verification workflows require transferring large data buffers (e.g., 4KB firmware pages) between protection domains for hashing and integrity checking. Currently, pw_kernel uses message-based IPC copying, which introduces unnecessary overhead:

• Throughput bottleneck: Page data must be copied from the caller's buffer into the IPC message for each transfer, with overhead proportional to page size
• Latency: Per-message copying overhead accumulates over many pages
• Not suitable for streaming: Hashing large firmware images page-by-page for measurement

[rusty1968]

@konkers , can you give your feedback?

[konkers]

This is a typical use case for shared memory. That's definitely on the burdock/cassava roadmap. Are you noticing specific performance issues w/ copying or is this a proactive request?

[rusty1968]

We absolutely need to share the non-cacheable memory between the measurement process and the crypto driver in this scenario. If we don't have it we would have to co-locate the measurement process and hash accelerator in the same process. Scoped lease is better for this because the sharing expires at the end of each transaction.