In uprobes_i386.c and uprobes_x86_64.c, we spend quite a bit of code on refusing to probe certain types of instructions, such as undefined opcodes and other instructions that are likely to yield SIGILL when single-stepped in user mode. There's also some paranoia about how instruction prefixes and such might affect our decisions about which instructions need to be treated specially when single-stepping them out of line. We could probably lighten up here. One disadvantage of the paranoia is that tests that try to probe (say) EVERY instruction in a .o file will choke on things like hlt. For illegal instructions, if the SIGILL kills the task before the SIGTRAP from the single-step is reported to us -- we need to test this -- then we should be OK. Testing suggests that we know how to handle tasks that die of other causes during probepoint processing. And I think we've successfully identified instructions that need special attention during SSOL. It's probably just a matter of testing each currently banned instruction type to verify that it doesn't present an unforeseen problem.
uprobes/uprobes_i386.c has been brought back into sync with the (more tolerant) uprobes2/uprobes_x86.c. A couple of changes were ad hoc to address specific test failures. Commit af6b060 completes the resync. We still haven't done any exhaustive testing of the whole instruction set.
*** Bug 10324 has been marked as a duplicate of this bug. ***
old rhel5 utrace/uprobes is unlikely to be perfected