[Frodocz]

Detailed descriptions will be shown in #1089.

ff_init() passes the global dpdk_argv[] straight to rte_eal_init(). DPDK's argument parser overwrites one slot with the program name at the end of parsing (lib/eal/linux/eal.c, eal_parse_args()):

if (optind >= 0)
    argv[optind-1] = prgname;

All F-Stack EAL arguments are options, so optind == argc and dpdk_argv[argc-1] now points to the same string as dpdk_argv[0]. With the default config, gdb after rte_eal_init() shows:

dpdk_argc = 4
dpdk_argv[0] = 0x555557737650 "f-stack"
dpdk_argv[1] = 0x5555577450d0 "-c1"
dpdk_argv[2] = 0x555557742780 "-n4"
dpdk_argv[3] = 0x555557737650 "f-stack"     <- was "--proc-type=auto"

ff_unload_config() (new in 2.0) frees every dpdk_argv[i], so "f-stack" is freed twice. The original dpdk_argv[argc-1] string leaks, which is the 17-byte dpdk_args_setup leak reported in #1089 on master. Any config hits this, because -c/-n are always generated.

The fix gives EAL a scratch copy of the pointer array, so dpdk_argv[] keeps the pointers ff_unload_config() owns. With the patch, the same repro under ASan runs to the end of ff_run() with no double free, and the dpdk_args_setup leak is gone.