zhhyu7 opened a new pull request, #22039:
URL: https://github.com/apache/nuttx/pull/22039
## Summary
Commit 7f04498f157 ("drivers/net/vlan: strip vlan devices from their
reliance on real device lock") removed netdev_lock() from around the receive
loop in netdev_upper_rxpoll_work(), so
that a VLAN-tagged frame could be dispatched to the VLAN device without
holding the real device lock. The non-VLAN branch of the same loop lost the
lock as well.
That branch still needs it. netdev_upper_input() documents "Called with the
network device locked", and netdev_upper_txavail_work() continues to take the
lock. With only one side locked,
receive and transmit race on dev->d_iob. devif_send() prepares dev->d_iob
through netdev_iob_prepare() and then copies into it with iob_trycopyin(); an
unlocked receive can replace or
release that IOB in between, and the copy gets a NULL chain:
Assertion failed : at file: iob/iob_copyin.c:69 task: hpwork
This takes the lock around each non-VLAN frame instead of around the whole
loop. netdev_upper_input() gets its lock back and is mutually exclusive with
the TX path again, while the VLAN
dispatch still runs without the real device lock, so the real-time
improvement of the original change is kept.
Reported-by: raiden00pl in
https://github.com/apache/nuttx/pull/21931#issuecomment-6977937414
## Impact
Fixes a crash on the receive path of any driver using the netdev upper half,
whenever TX runs concurrently with RX. CONFIG_NET_VLAN does not have to be
enabled to be affected.
The lock is now taken and released once per received frame rather than once
per poll batch. This is the rmutex fast path when uncontended, and the previous
code held the lock for the wholebatch anyway, so throughput is not expected to
change. Measured 4 x 256 KiB of concurrent TCP echo with no regression.
No API or configuration change.
## Testing
Host: x86_64 Linux. Configuration: sim:dynconns (flat build) with the TAP
device bridged to the host bridge nuttx0 (10.0.1.1/24); the device is
10.0.1.2/24. For the test I enabled
DEBUG_ASSERTIONS, EXAMPLES_TCPECHO, SIM_NET_BRIDGE, FS_PROCFS and
NETINIT_THREAD, disabled NETDEV_LATEINIT, and raised the IOB pool to
IOB_NBUFFERS=256 / IOB_BUFSIZE=1536. CONFIG_NET_VLAN was left disabled, which
shows the bug is not specific to VLAN.
The reported case, 4 concurrent TCP echo connections of 256 KiB each.
Before, on 7f04498f157:
conn0: sent=2048 recv=0 err=TimeoutError('timed out')
conn1: sent=2048 recv=64 err=TimeoutError('timed out')
conn2: sent=2048 recv=0 err=TimeoutError('timed out')
conn3: sent=2048 recv=0 err=TimeoutError('timed out')
dump_assert_info: Assertion failed : at file: iob/iob_copyin.c:69 task:
hpwork process: Kernel 0x400034e0
After:
conn0: sent=262144 recv=262144 err=None
conn1: sent=262144 recv=262144 err=None
conn2: sent=262144 recv=262144 err=None
conn3: sent=262144 recv=262144 err=None
The crash needs far less than that. A single 64-byte TCP connection is
enough, while ICMP alone never triggers it, which matches the analysis: TCP has
a send path that can sit between
netdev_iob_prepare() and iob_trycopyin(), and a ping does not.
Before, on 7f04498f157:
=== only ping, no TCP ===
assert_count_after_ping=0
=== one TCP echo of 64 bytes ===
sent=64 recv=64 match=True
assert_count_after_tcp=1
After, same two steps:
assert_count_after_ping=0
sent=64 recv=64 match=True
assert_count_after_tcp=0
Note that the echo itself succeeds before the assertion fires, so the crash
is not a data path failure.
./tools/checkpatch.sh -c -u -f drivers/net/netdev_upperhalf.c passes. Build
and link succeed.
--
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]