casaroli opened a new pull request, #21906:
URL: https://github.com/apache/nuttx/pull/21906

   ## Summary
   
   `pkt_close()` freed the read-ahead and error-ahead queues of a packet socket 
before the connection left the active list. The input path can still queue 
frames on the connection after that point. `pkt_free()` then released the 
connection with these IOBs still queued, and nothing freed them.
   
   A frame can come in at two points. With `CONFIG_NET_PKT_WRITE_BUFFERS`, 
`pkt_close()` waits for the write queue to drain with the device unlocked: a 
frame that comes in during this wait is queued, for example a sent frame that 
comes back on `lo`. Also, a frame can come in between `conn_dev_unlock()` and 
`pkt_free()`.
   
   This PR frees the queues in `pkt_free()` after `dq_rem()`, with the 
connection list locked, as `can_conn_uninit()`, `icmp_free()` and `udp_free()` 
do. `pkt_in_()` queues frames with the same lock held, so no frame can come in 
after that point.
   
   ## Impact
   
   Closing a packet socket no longer loses IOBs. Before, each such close made 
the IOB pool of the network stack smaller for good. No API change. No 
Documentation change (`Documentation/os/networking/pkt.rst` does not describe 
this). The code is the same in all build modes and on all arches.
   
   ## Testing
   
   Two test programs, each with an `ETH_P_ALL` packet socket bound to `lo` that 
receives IPv4 packets to 127.0.0.1. Each compares the free IOB count of 
`/proc/iobinfo` at the start and at the end.
   
   - `pktclose.c`: 4 rounds of open, send 2 packets, close. Case 1 at normal 
priority (the frames are queued before `close()`). Case 2 above the work queue 
priority, so `close()` waits for the write queue while the frames come back to 
the socket.
   - `pktrace.c`: open and close the socket in a loop for 3 s, while a thread 
at priority 200 sends a UDP datagram to 127.0.0.1 at each clock tick. A frame 
that comes in during `close()` is lost on master.
   
   | Configuration | master 46066440966 | this PR |
   |---|---|---|
   | qemu-armv8a:knsh (kernel), write buffers | pktclose case 2: 16 IOBs lost, 
FAIL | pktclose: 0 lost, PASS |
   | qemu-armv8a:nsh (flat), write buffers | pktclose case 2: 16 IOBs lost, 
FAIL | pktclose: 0 lost, PASS |
   | qemu-armv8a:nsh (flat), no write buffers | pktclose: PASS; pktrace: 8 IOBs 
lost after 136896 rounds, FAIL | pktclose: PASS; pktrace: 0 lost after 3196984 
rounds, PASS |
   | rv-virt:knsh64 (kernel), write buffers | not run | pktclose and pktrace: 
PASS |
   
   ostest exits with status 0 on qemu-armv8a:knsh, on qemu-armv8a:nsh (with and 
without write buffers) and on rv-virt:knsh64. All runs are under QEMU (arm64 
with HVF, rv-virt with TCG), one CPU. Options added to the defconfigs: NET, 
NET_IPv4, NET_LOOPBACK, NET_UDP, NET_ICMP, NET_PKT, NET_PKT_WRITE_BUFFERS (on 
or off), DEBUG_ASSERTIONS.
   
   checkpatch: `tools/checkpatch.sh -c -u -m -g` passes.
   


-- 
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]

Reply via email to