> Hi Gordon, Hi Rob,
Thanks for having a look at this stuff. > I am mostly happy with your changes; I think they're a > useful addition to a test suite that gets intermittent > love :-) > > One thing - the special tests used to copy themselves to > the target and run from there, and you have changed it to > at least not always do that. I understand why we need a > way to avoid executing from the target remote filesystem, > but executing that way has at least a time or two shown > off bugs, mostly in the client, because of the way exec() > manages mappings of the binary. I would prefer that this > continue to be the default mode of operation to maintain > this coverage, and to have an option to have the special > tests do a chdir() instead. This is not huge, but I'd > rather not lose the regression coverage. > > Rob T I could make that change to the "special" tests optional, but I'm not sure that would really be an improvement. The changes to the "special" tests to run them from where they are instead of copying into the mount point was for two reasons: (1) allow running them on an FS with "nomap", and (2) make the tests cover just the specific features each was designed to exercise. It appeared to me that covering exec/mmap in those tests was an accident of implementation. Further, if there are failures in what was being tested by the "special" tests, the practice of running the tests in the mount would only serve to complicate the debugging efforts because much of the access via the mount would have nothing to do with what the tests are really trying to cover. One could interpret the change in the "special" tests as a reduction of coverage (because it no longer covers exec) but the test suite as a whole does appear to cover the mmap/exec cases reasonably well in the "general" tests. What I'd prefer to do with this is to augment the "general" tests (if necessary) to cover whatever we think may be no longer adequately covered after the changes in "special". Would you be OK with that plan? Can you think of any specific exec/mmap cases not covered by the current "general" tests? (i.e. the test failure you mentioned you had seen previously?) Thanks, Gordon
