Hello, I'm using the following version of GCC: artpol@artpol-ThinkPad-T430 ~ $ gcc --version gcc (Ubuntu 4.8.4-2ubuntu1~14.04.1) 4.8.4 Copyright (C) 2013 Free Software Foundation, Inc. This is free software; see the source for copying conditions. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
I'm facing the problem with the slurm compilation: https://github.com/SchedMD/slurm I'm building the master branch. And the problem I get is related to the following construction: https://github.com/SchedMD/slurm/blob/master/src/slurmd/slurmd/req.c#L6171:L6172 The problem I see is that "connect" returns "+1" instead of "-1". And I can't reproduce this with the standalone code: https://github.com/artpol84/poc/blob/master/linux-internals/syscalls/connect/cli_c1.c#L37 Here is what GDB shows for those 2 codes: SLURM's req.c: 0x43e966 <_rpc_forward_data+326> lea 0x2(%rax),%edx 0x43e969 <_rpc_forward_data+329> lea -0x80(%rbp),%rcx 0x43e96d <_rpc_forward_data+333> mov -0xc4(%rbp),%eax 0x43e973 <_rpc_forward_data+339> mov %rcx,%rsi 0x43e976 <_rpc_forward_data+342> mov %eax,%edi 0x43e978 <_rpc_forward_data+344> callq 0x4287e0 <connect@plt> # eax == -1 0x43e97d <_rpc_forward_data+349> shr $0x1f,%eax # eax == 1 0x43e980 <_rpc_forward_data+352> movzbl %al,%eax # eax == 1 0x43e983 <_rpc_forward_data+355> mov %eax,-0xc0(%rbp) # (gdb) p $rbp # $17 = (void *) 0x7fcb8e15ee60 # (gdb) p &rc # $15 = (int *) 0x7fcb8e15eda0 # $rbp - 0xC0 == &rc 0x43e989 <_rpc_forward_data+361> cmpl $0x0,-0xc0(%rbp) # if comparison Standalone program: 0x400863 <usock_connect+198> mov %rcx,%rsi 0x400866 <usock_connect+201> mov %eax,%edi 0x400868 <usock_connect+203> callq 0x400690 <connect@plt> # eax == -1 0x40086d <usock_connect+208> mov %eax,-0x94(%rbp) # (gdb) p &rc # $1 = (int *) 0x7fffffffdeac # (gdb) p $rbp - 0x94 # $2 = (void *) 0x7fffffffdeac # => we just move ( $eax == -1 ) to the rc. 0x400873 <usock_connect+214> cmpl $0x0,-0x94(%rbp) 0x40087a <usock_connect+221> jns 0x400888 <usock_connect+235> So for some reasons the code issued by compiler for equal codes is different. The compilation flags used was: req.c: gcc -DHAVE_CONFIG_H -I. -I../../.. -I../../../slurm -DLIBSLURM_SO=\"/lxc-data/sandbox/slurm/lib/libslurm.so\" -I../../.. -g -O2 -pthread -Wall -g -O0 -fno-strict-aliasing -MT req.o -MD -MP -MF .deps/req.Tpo -c -o req.o req.c standalone (was trying to use the same flags): gcc -g -O2 -pthread -Wall -g -O0 -fno-strict-aliasing -c -o cli_c1.o cli_c1.c FIX that changes the situation in SLURM: if I apply attached patch to the SLURM sources that shightly changes C construction: - while ((rc = connect(fd, (struct sockaddr *)&sa, SUN_LEN(&sa)) < 0) && - (errno == EINTR)); + + do { + rc = connect(fd, (struct sockaddr *)&sa, SUN_LEN(&sa)); + } while( (rc < 0) && (errno == EINTR) ); I'm getting correct code: 0x43e971 <_rpc_forward_data+337> mov %eax,%edi 0x43e973 <_rpc_forward_data+339> callq 0x4287e0 <connect@plt> 0x43e978 <_rpc_forward_data+344> mov %eax,-0xc0(%rbp) 0x43e97e <_rpc_forward_data+350> cmpl $0x0,-0xc0(%rbp) ---- Best regards, Artem Y. Polyakov HPC Engineer, Mellanox Ltd.
