Sarthak-Pandey opened a new pull request, #20627:
URL: https://github.com/apache/nuttx/pull/20627
## Summary
This PR implements complete TCP urgent (Out-of-Band / OOB) data support for
the Apache NuttX networking stack under `CONFIG_NET_TCPURGDATA`, resolving
long-standing issue #2327 and addressing critical design review bottlenecks
across RX, TX, and socket API layers.
When `CONFIG_NET_TCPURGDATA` is disabled (`n`), the build output and memory
footprint remain 100% unchanged.
---
## Key Changes & Architecture
### 1. Data Structures & Semantics
- **Standard Semantics:** Adopts BSD/Linux standard semantics
(`tcp_stdurg=0`), where the urgent pointer points one byte past the urgent
byte, and the single OOB byte is `payload[urgp - 1]`.
- **SO_OOBINLINE:** When `SO_OOBINLINE` is enabled, urgent data remains
in-line in the normal byte stream without being extracted.
- **Bounded Buffering:** Implemented BSD single-byte in-place overwrite
policy in `conn->urgreadahead` to prevent IOB pool exhaustion on high-frequency
URG floods.
### 2. RX Path (`net/tcp/tcp_input.c`)
- Extracted the single urgent byte into `conn->urgreadahead` and compacted
remaining payload in-place using `memmove()` without distorting stream sequence
numbers.
- **Allocation Failure Safety:** If IOB allocation fails, the packet is
dropped without advancing `rcvseq`, forcing the peer TCP stack to retransmit
(avoiding silent data loss).
- **Pure-Urgent ACKs:** Added `TCP_SNDACK` trigger when urgent data is
consumed so that pure-urgent single-byte segments (`d_len == 0`) receive an
immediate ACK.
### 3. Socket API & Polling (`net/tcp/tcp_recvfrom.c`,
`net/tcp/tcp_netpoll.c`, `net/devif/devif.h`)
- Implemented `tcp_recv_oob()` for `recv(..., MSG_OOB)`.
- **Safe Waiter Lifecycle:** Replaced fragile connection-level semaphore
(`urg_sem`) with the standard `devif_callback` pattern (`TCP_URGDATA |
TCP_DISCONN_EVENTS | TCP_RXCLOSE`). This eliminates use-after-free crashes on
`close()`, hangs on remote RST / abort / timeout, and stale semaphore counts.
- **Reliable `POLLPRI` Wakeup:** Defined conflict-free event flag
`TCP_URGDATA (1 << 25)` in `devif.h` to immediately awaken threads blocked in
`poll(POLLPRI)` even when payload length drops to 0.
### 4. TX Path (`net/tcp/tcp_send.c`, `net/tcp/tcp_send_buffered.c`,
`net/tcp/tcp_send_unbuffered.c`)
- Removed inbound URG flag inheritance in `tcp_send()` so that plain ACKs
and replies never leak incoming `TCP_URG` or stale `urgp`.
- In-order buffer queueing via `sq_addlast` prevents wire reordering of
prior normal data.
- Per-segment urgent pointer calculation: `conn->urg_offset = (wrb->wb_urg -
TCP_WBSENT(wrb))`.
---
## ❓ Questions for Maintainers / Reviewers
Before merging, I would love feedback from the maintainers on the following
points:
1. **Urgent Buffering Policy:**
- Currently, we follow the BSD/Linux default policy of keeping at most
**1 unread urgent byte** per connection (new urgent byte overwrites the
previous unread byte in-place, consuming 0 additional IOBs).
- *Question for maintainers:* Is this single-byte overwrite policy
preferred, or would you like me to introduce an optional Kconfig knob (e.g.
`CONFIG_NET_TCP_URG_BUFCAP`, defaulting to 1 but allowing a FIFO queue of up to
N bytes before trimming the oldest)?
2. **SO_OOBINLINE & `sockatmark()`:**
- With `SO_OOBINLINE` active, urgent bytes stay inline in the normal read
stream. Should we also introduce `sockatmark()` support in this PR, or track it
as a subsequent follow-up?
3. **Nagle Bypass on Urgent Buffers:**
- In `tcp_send_buffered.c`, urgent buffers are queued in-order. Would
maintainers prefer an explicit Nagle bypass (forcing immediate segment flush
with `TCP_PSH`) when `wb_urg > 0` is present?
---
## Verification
- Clean run on `tools/nxstyle` across all modified files.
- Build verified with:
- `CONFIG_NET_TCPURGDATA=n` (verified zero diff in binary/data layout)
- `CONFIG_NET_TCPURGDATA=y` with `CONFIG_NET_TCP_WRITE_BUFFERS=y`
- `CONFIG_NET_TCPURGDATA=y` with `CONFIG_NET_TCP_WRITE_BUFFERS=n`
- Verified no remaining `tcp->flags & TCP_URG` reads from inbound headers.
- Verified no remaining `sq_addfirst` for urgent buffers.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
