Hi,
We see a crash in the dpdk packet distributor library when the packets are
assigned used defined flow ids. I was able to narrow down this issue to
incorrect worker ids returned by the sse implementation of find_match_vec().
Debug:
Non zero flow ids picked up from each mbuf in this burst
(gdb) p flows
$63 = {9521, 9521, 9521, 9521, 9521, 9521, 9521, 9521}
There are eight workers:
(gdb) p d->num_workers
$64 = 8
And eight packets being processed in this burst
(gdb) p pkts
$65 = 8
A call using vec instructions returns incorrect worker ids:
(gdb) call find_match_vec(d, &flows[0], &matches[0])
(gdb) p matches
$66 = {9, 9, 9, 9, 9, 9, 9, 9}
Where as, a call to the scalar implementation returns workers ids up to 8.
A comment in rte_distributor_process_v1705() says the matches array now
contain the intended worker ID (+1). So it makes sense to have worker ids
up to eight. (0-7)+1
(gdb) call find_match_scalar(d, &flows[0], &matches[0])
(gdb) p matches
$67 = {1, 1, 1, 0, 8, 0, 0, 0}
Is this a bug in the sse implementation of the find match function?
The function SEG faults when trying to access non-existent workers backlog
structure:
(gdb) p d->backlog[matches[j]-1]
$76 = {start = 0, count = 1, pkts = {0, 0, 0, 0, 0, 0, 0, 0}, tags = 0x0}
Regards,
Suryanathan
