[dpdk-dev] Possible bug in eal_pci pci_scan_one

[Stephen Hemminger]

On Mon, 6 Oct 2014 02:13:44 -0700
Matthew Hall <mhall at mhcomputing.net> wrote:

> Hi Guys,
> 
> I'm doing my development on kind of a cheap machine with no NUMA support... 
> but several years ago I used DPDK to build a NUMA box that could do 40 gbits 
> bidirectional L4-L7 stateful traffic replay.
> 
> So given the past experiences I had before, I wanted to clean the code up so 
> it'd work well if some crazy guy tried my code on one of these huge boxes, 
> too, but then I ran into some weird issues.
> 
> 1) When I call rte_eth_dev_socket_id() I get back -1. But the call can return 
> -1 if the port_id is bogus or if pci_scan_one didn't get a numa_node (because 
> you're on a non-NUMA box for example).
> 
> int rte_eth_dev_socket_id(uint8_t port_id)
> {
>         if (port_id >= nb_ports)
>                 return -1;
>         return rte_eth_devices[port_id].pci_dev->numa_node;
> }
> 
> So you couldn't tell the different between non-NUMA or a bad port value, etc.
> 
> 2) The code's behavior and comments disagree with one another. In the 
> pci_scan_one function, there's this code:
> 
> /* get numa node */
> snprintf(filename, sizeof(filename), "%s/numa_node",
>          dirname);
> if (access(filename, R_OK) != 0) {
>         /* if no NUMA support just set node to 0 */
>         dev->numa_node = -1;
> } else {
>         if (eal_parse_sysfs_value(filename, &tmp) < 0) {
>                 free(dev);
>                 return -1;
>         }
>         dev->numa_node = tmp;
> }
> 
> It says, just use NUMA node 0 if there is no NUMA support. But then proceeds 
> to set the value to -1 in disagreement with the comment, and also stomping on 
> the other meaning for -1 in the higher function rte_eth_dev_socket_id.
> 
> 3) In conclusion, it seems like some stuff is missing... first there needs to 
> be a function that will tell you the number of NUMA nodes present on the box 
> so you can create the right number of mbuf_pools, but I couldn't find that function.
> 
> Then if you have the function, you can do some magic and shuffle the NICs 
> around to get them hooked to a core on the same NUMA, and the mbuf_pool on the 
> same NUMA.
> 
> When NUMA is not present, can we return 0 instead of -1, or return a specific 
> error code that the client can use to know he should just use Socket 0? Right 
> now I can't tell apart any potential errors or weird values from correct 
> values.
> 
> 4) I'm willing to help make and test some patches... but first I want to 
> understand what is happening with these funny functions before doing things 
> blindly.
> 
> Thanks,
> Matthew.

The code is fairly consistent in returning -1 for cases of not a NUMA socket,
bogus port value. It is interpreted as SOCKET_ID_ANY in several places.
The examples mostly check for -1 and use socket 0 as a fallback.
Probably not worth introducing more return values and breaking existing
applications.