FS-Tee Grate.

[stupendoussuperpowers]

No description provided.

[rennergade]

should just be 'fs-tee' now

[rennergade]

Left a handful of comments

[rennergade]

Seems like theres a merge conflict. Some thoughts I have:

• The tee dispatcher runs the primary syscall first, then runs the secondary syscall using the same
  guest pointers and buffers:

  rust-grates/fs-tee-grate/src/handlers/tee.rs:84

  For read-like syscalls, this means the return value comes from the primary path, but the bytes left
  in target memory may come from the secondary path. Examples: read, pread, readv, stat, getdents,
  readlink, etc.

  fs-tee should either avoid running secondary copyout syscalls against the target’s real output
  buffers, or copy/restore the primary output before returning.

• pipe()/pipe2() fallback calls SYS_OPEN and does not record pipe FD mappings

  In tee_pipe and tee_pipe2, the fallback path uses SYS_OPEN:

  rust-grates/fs-tee-grate/src/handlers/tee.rs:334
  rust-grates/fs-tee-grate/src/handlers/tee.rs:613

  That should be SYS_PIPE / SYS_PIPE2. Also, the handlers do not record the two returned pipe FDs, so
  later reads/writes on those pipe descriptors cannot be translated to the secondary path.

  This breaks tee behavior for programs that use pipe or pipe2 unless the secondary grate happens to
  provide a matching alternate handler.

• Secondary handler lookup picks the first registration, not the topmost one

  get_interposition_request uses .find():

  rust-grates/fs-tee-grate/src/utils.rs:33

  If multiple grates in the secondary stack register the same syscall, this selects the first one that
  registered. In a stacked grate chain, the later/topmost registration should win. This should likely
  use .rfind(), matching the intended handler stack order.

• file-backed mmap is mirrored without translating the fd

  mmap is registered as a path/no-FD-translation handler:

  rust-grates/fs-tee-grate/src/handlers/tee.rs:191

  But file-backed mmap uses arg5 as the fd. If primary and secondary fd numbers differ, the secondary
  mmap runs against the wrong fd or fails. mmap needs special handling: skip fd translation for
  anonymous mappings, but translate arg5 for file-backed mappings.

[stupendoussuperpowers]

@rennergade, addressed all these. Tests still pass too. Feel free to review again.

[stupendoussuperpowers]

Added a fdtables::close in the tee_close handler. Along with error checks before record_fd_pair and in the fork handler.

[rennergade]

The exerciser calls dup/dup2/dup3, but the test only checks that the target exits successfully. It does not assert that the secondary path saw successful mirrored results or that fs-tee-secondary.log has matching success values. That means the fd translation bug above can pass if the primary path works while secondary calls fail/log -1.

[stupendoussuperpowers]

Do you have any proposals on what a test for the secondary paths should look like? Those values are never really returned anywhere.

Checking the log file seems to be unreliable.

I don't think this is limited to dup* either, pretty much all the tests above are only checking to make sure the primary paths are not hampered.

[rennergade]

put in some comments related to the dup family and a small fix to the test runner that should make it more reliable.