The function wordexp() fails to properly handle the WRDE_NOCMD flag when processing arithmetic inputs in the form of "$((... ``))" where "..." can be anything valid. The backticks in the arithmetic epxression are evaluated by in a shell even if WRDE_NOCMD forbade command substitution. This allows an attacker to attempt to pass dangerous commands via constructs of the above form, and bypass the WRDE_NOCMD flag. This patch fixes this by checking for WRDE_NOCMD in exec_comm(), the only place that can execute a shell. All other checks for WRDE_NOCMD are superfluous and removed. We expand the testsuite and add 3 new regression tests of roughly the same form but with a couple of nested levels. On top of the 3 new tests we add fork validation to the WRDE_NOCMD testing. If any forks are detected during the execution of a wordexp() call with WRDE_NOCMD, the test is marked as failed. This is slightly heuristic since vfork might be used in the future, but it provides a higher level of assurance that no shells were executed as part of command substitution with WRDE_NOCMD in effect. In addition it doesn't require libpthread or libdl, instead we use the public implementation namespace function __register_atfork (already part of the public ABI for libpthread). Tested on x86_64 with no regressions. Diff: https://sourceware.org/git/gitweb.cgi?p=glibc.git;a=commitdiff;h=a39208bd7fb76c1b01c127b4c61f9bfd915bfe7c