fix(igc): buffer rx packets on irq

[ytakano]

This is PR4 in the igc incremental stack (stacked on top of PR3 / igc_observability; PR1–3 are already merged).

PR4 changes (this commit)

Decouples RX packet delivery from immediate upper-layer consumption by
introducing a software ring queue drained during interrupt handling.

Key modifications to awkernel_drivers/src/pcie/intel/igc.rs:

• Adds read_queue: RingQ<EtherFrameBuf> (32-packet capacity) to Rx
• Replaces igc_recv() with igc_rx_recv(que_id, rx: &mut Rx) that
• recv() fast-path pops from read_queue; falls back to igc_rx_recv()
• intr() calls igc_rx_recv() on RX IRQ, buffering packets before upper-layer consumption
• Resets read_queue in igc_allocate_queues(), igc_setup_rx_desc_ring(), and igc_stop()

applications/tests/test_network adjusted: INTERFACE_ID → 1, addresses → 192.168.100.x, tests narrowed to tcp_listen_test + new udp_recv_test.

PR4 review follow-up (also in this commit)

Addressed Copilot review feedback:

• Fixed no-op self.dump() in igc_up(): captures and logs the returned String via log::debug! (regression introduced during PR3 refactoring when dump() was changed to return String instead of logging directly)
• Changed let _ = rx.read_queue.push(packet) to increment dropped_pkts on failure, making any unexpected push failure observable via the existing drop counter (the while !is_full() guard makes this path unreachable in practice)
• Removed "via log::debug!" from debug_dump_interface() doc comment; the log level is device-defined
• Updated (netdump if) help text to (netdump id) to avoid the keyword-like if

Not adopted: kept warn! in the default NetDevice::debug_dump() (PR3 design decision — operators should see a clear message when calling netdump on a device that has not implemented the hook); did not alter the PR
description for debug_dump/netdump additions (those are PR3 changes visible because PR4 is stacked on PR3).

Deferred to later PRs

• TX reclaim IRQ path (PR5+): can_send() remains a reclaiming gate
• Multi-queue support (PR5)
• TX checksum offload (PR6)

Testing

• make clippy — clean
• make x86_64 RELEASE=1 — successful
• make check_x86_64 — passed
• Real hardware: IGC ports initialized; "link status changed: Up (Full Duplex)" at 1000 Mbps confirmed

[Copilot]

Pull request overview

This PR aims to improve the Intel IGC driver RX path by buffering received packets into a software ring queue during IRQ handling, decoupling descriptor consumption from immediate upper-layer reads. It also introduces an on-demand “netdump” facility by adding a NetDevice::debug_dump() hook, a debug_dump_interface() helper, and a corresponding shell command.

Changes:

• Add an RX-side software ring queue (RingQ<EtherFrameBuf>) to IGC and drain completed RX descriptors into it from the IRQ path; update recv() to pop from this queue first.
• Add a generic NetDevice::debug_dump() API + awkernel_lib::net::debug_dump_interface() helper, and wire a shell (netdump <interface_id>) command to invoke it.
• Refactor IgcInner::dump() to return a String so callers decide how/when to emit logs.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
awkernel_lib/src/net/net_device.rs	Adds NetDevice::debug_dump() default implementation (new API surface).
awkernel_lib/src/net.rs	Adds debug_dump_interface() helper that looks up an interface and invokes debug_dump() outside the NET_MANAGER lock.
awkernel_drivers/src/pcie/intel/igc.rs	Implements RX buffering via RingQ, calls RX drain from IRQ, updates recv(), and changes dump() to return String.
applications/awkernel_shell/src/lib.rs	Registers a new BLisp-exported (netdump ...) command and FFI to call debug_dump_interface().
applications/awkernel_shell/Cargo.toml	Adds num-bigint / num-traits dependencies to parse BLisp Int into u64 for netdump.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[atsushi421]

Using ? here short-circuits the IMS/EIMS re-enable writes at the end of intr(). Since the ICR read has already ACKed the interrupt at the device, an error here can leave LSC/per-queue interrupts un-rearmed and silently drop subsequent IRQs. Capture the rx/tx results, still perform the rearm writes, and only then propagate the error.

[atsushi421]

The old igc_recv() had a defensive if que_id != 0 { return Ok(None); }, but igc_rx_recv() drops that guard while the PR body explicitly defers multi-queue to PR5. If irqs_to_queues ever yields a non-zero que_id before PR5 lands, this will touch ring state that is not yet wired up. Reinstate the guard (or a debug_assert_eq!(que_id, 0)) until PR5 enables multi-queue.

[atsushi421]

A full memory barrier on every descriptor in the drain loop is more expensive than necessary. The barrier only needs to pair once with the device's DMA write-back; subsequent descriptors are picked up on later iterations or later IRQs. Move membar_sync() out of the loop (or replace the status-word read with an acquire-ordered load).

[atsushi421]

recv() locks rx, pops, drops the lock, then re-locks it to call igc_rx_recv() and pop again. The two acquisitions are pure overhead and also open a window where intr() can mutate read_queue between them. Collapse into a single critical section: lock once, try pop(), and only call igc_rx_recv() (and pop again) if pop returned None.

[atsushi421]

The PR description states this was changed from (netdump if) to (netdump id) to avoid the keyword-like if, but the diff still says if. The change appears to have been lost; please apply it (and consider clarifying that the argument is the numeric interface_id).

[atsushi421]

The PR description says this was changed to increment dropped_pkts on push failure, but the diff still uses let _ = rx.read_queue.push(packet);. The promised observability is missing — any unexpected push failure remains invisible. Replace with if rx.read_queue.push(packet).is_err() { rx.dropped_pkts = rx.dropped_pkts.saturating_add(1); break; } (the break also avoids spinning until the next is_full() check).