Hi Andrew, You observation is right. I was running vpp_lite with a buffer size of 512. As you mentioned, defining PING_MAXIMUM_DATA_SIZE conditionally should work.
I have opened a jira ticket for this: https://jira.fd.io/browse/VPP-621 Thanks, -nagp On Fri, Jan 27, 2017 at 7:23 PM, Andrew 👽 Yourtchenko <ayour...@gmail.com> wrote: > Hello, > > > On 27 Jan 2017, at 04:12, Nagaprabhanjan Bellaru <nagp.li...@gmail.com> > wrote: > > > > Hi, > > > > I am not sure if the ping debug CLI is being actively used, but the > function "init_icmp46_echo_request" goes ahead and writes 2000 bytes into > the vlib_buffer corrupting the surrounding memory area. After 3-4 pings, > vpp always crashes. > > Could you please tell a bit more about the setup (which hypervisor, which > make platform - vpp/vpp_lite, which ping etc) ? I had fixed a bug with > processing a vector of replies (change 4844) - might be worth verifying you > do not see the symptoms of that. > > Now, some thoughts below, please feel free to correct me if you find an > error in the below logic. > > My understanding is we would get the 2048 bytes of max data size in DPDK > case when allocating the buffer. Minus 20 bytes for IPv4 header, minus 4 > bytes for common ICMP header, minus 4 bytes for echo ID/seq, minus 8 bytes > for the timestamp. That should give 2012 bytes of free space for data in > IPv4 case, so for the IPv4 ping it should not overrun. > > The problem is of course in IPv6 case we are a few bytes short, and > likewise in the vpp_lite case the VLIB_BUFFER_DATA_SIZE is 512, and we > will overrun that with the static value of 2000. > > So the definition for PING_MAXIMUM_DATA_SIZE would need to depend on that > define with appropriate subtractions... or, even better, I suppose, is to > use VLIB_BUFFER_DEFAULT_FREE_LIST_BYTES since that looks to me is the > define determining the max data size for a buffer. > > Maybe an even better option could be to just fill a vector and then use > the vlib_buffer_add_data() with zero buffer index, and let it allocate and > set up everything behind the scenes. > > What do you think ? > > > > > Instead of copying sizeof(icmp_echo_request->data) which is 2000, it we > copy just data_len to the buffer, it should be fine? > > That would make the problem conditional on the CLI input for data size, so > I would not say it is a much better outcome... > > --a > > > > > > Thanks, > > -nagp > > _______________________________________________ > > vpp-dev mailing list > > vpp-dev@lists.fd.io > > https://lists.fd.io/mailman/listinfo/vpp-dev >
_______________________________________________ vpp-dev mailing list vpp-dev@lists.fd.io https://lists.fd.io/mailman/listinfo/vpp-dev
