* [dpdk-users] Possible memory corruption due to incorrect DMA shutdown
@ 2016-03-11 18:41 George Prekas
2016-09-21 13:09 ` George Prekas
0 siblings, 1 reply; 2+ messages in thread
From: George Prekas @ 2016-03-11 18:41 UTC (permalink / raw)
To: users
Hi. I've been using DPDK for a research project
(https://www.usenix.org/conference/osdi14/technical-sessions/presentation/belay)
for over 2 years and I'd like to report a behavior that puzzles me using
DPDK.
The behavior leads to memory corruption and is caused by the incorrect
shutdown of DMA. I can reproduce it after executing the following steps:
$ sudo modprobe uio
$ sudo insmod ./build/kmod/igb_uio.ko
$ sudo python ./tools/dpdk_nic_bind.py --bind=igb_uio 0000:42:00.1
$ sudo ./build/app/testpmd -- --forward-mode=icmpecho
Then I terminate the DPDK program (after populating the ARP cache of
another host on the local network). After this, I can send UDP packets
to the host and can observe their payload in host memory. Clearly,
network packets are arriving to the network card and are written to RAM
after DPDK has finished executing.
Am I doing something wrong? Is this behavior expected?
Regards,
George
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [dpdk-users] Possible memory corruption due to incorrect DMA shutdown
2016-03-11 18:41 [dpdk-users] Possible memory corruption due to incorrect DMA shutdown George Prekas
@ 2016-09-21 13:09 ` George Prekas
0 siblings, 0 replies; 2+ messages in thread
From: George Prekas @ 2016-09-21 13:09 UTC (permalink / raw)
To: users
Any updates on this?
I can consistently reproduce this behavior following these steps:
On host A do:
$ sudo modprobe uio
$ sudo insmod ./build/kmod/igb_uio.ko
$ sudo python ./tools/dpdk_nic_bind.py --bind=igb_uio 0000:42:00.1
$ sudo ./build/app/testpmd -- --forward-mode=icmpecho
From host B (which is on the same local network), do an arping to an
arbitrary IP address C.C.C.C (that is in the same network and it doesn't
belong to another host). DPDK will respond to any IP address. All you
want is to populate host B's ARP cache. Then terminate DPDK and run the
following Python program on host B:
import sys
import socket
A=sys.argv[3] * int(sys.argv[2])
for i in xrange(10000):
sock = socket.socket(socket.AF_INET,socket.SOCK_DGRAM)
sock.sendto(A, (sys.argv[1], 1))
as:
$ python go.py C.C.C.C 28 prekageo_was_here_
Then go back to host A and use the physical memory grep tool that you
can find here https://github.com/prekageo/allmemscan
$ make
$ sudo insmod allmem.ko
$ sudo ./allmemscan 28 prekageo_was_here
Optionally:
$ sudo dd if=/dev/allmem bs=4K skip=PAGE count=1 | hexdump -C | less
On 11/03/2016 19:41, George Prekas wrote:
> Hi. I've been using DPDK for a research project
> (https://www.usenix.org/conference/osdi14/technical-sessions/presentation/belay)
> for over 2 years and I'd like to report a behavior that puzzles me
> using DPDK.
>
> The behavior leads to memory corruption and is caused by the incorrect
> shutdown of DMA. I can reproduce it after executing the following steps:
>
> $ sudo modprobe uio
> $ sudo insmod ./build/kmod/igb_uio.ko
> $ sudo python ./tools/dpdk_nic_bind.py --bind=igb_uio 0000:42:00.1
> $ sudo ./build/app/testpmd -- --forward-mode=icmpecho
>
> Then I terminate the DPDK program (after populating the ARP cache of
> another host on the local network). After this, I can send UDP packets
> to the host and can observe their payload in host memory. Clearly,
> network packets are arriving to the network card and are written to
> RAM after DPDK has finished executing.
>
> Am I doing something wrong? Is this behavior expected?
>
> Regards,
> George
>
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2016-09-21 13:09 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2016-03-11 18:41 [dpdk-users] Possible memory corruption due to incorrect DMA shutdown George Prekas
2016-09-21 13:09 ` George Prekas
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).