[Bug tapsets/15219] syscall.exp failures on RHEL5, RHEL6, and rawhide

["dsmith at redhat dot com" <sourceware-bugzilla at sourceware dot org>]

http://sourceware.org/bugzilla/show_bug.cgi?id=15219

--- Comment #7 from David Smith <dsmith at redhat dot com> ---
Created attachment 7067
  --> http://sourceware.org/bugzilla/attachment.cgi?id=7067&action=edit
anti-nested syscall patch

Here's a new patch that attempts (and fails) to fix this problem a different
way.

Instead of changing the way we access memory, instead this patch tries to avoid
the nested system calls. To do this, in the 64-bit syscall .call and .return
probe, I added the following line:

    if (%{ _stp_is_compat_task() %}) next

So, if we're in a 64-bit syscall with a 32-bit process, skip the probe (and let
the 32-bit syscall handle it). (That line could be prettied up by using a
macro.)

As I mentioned above, the advantage this approach has is that it doesn't change
the way we access memory. The disadvantages are:

- Knowing when and where to add that line. The problem is that some 32-bit
syscalls are wrappers around the 64-bit syscall, and some 32-bit syscalls call
the same underlying function that the 64-bit syscall does (after massaging the
arguments a bit). So, in the 1st case (the wrapper case), we need the probe
skip line, but in the 2nd case we don't.  I used the syscall.exp test case to
know where to put that line when developing the patch, but the syscall.exp test
case doesn't test every system call.

- Whether we need that line or not can change with kernels. For example, the
syscall.rt_sigprocmask probe should be skipped on RHEL[56] kernels, but not on
rawhide.

So, some more work will be required here.

-- 
You are receiving this mail because:
You are the assignee for the bug.