rzezeski
commented
2 years ago So I initially thought this was because opte is not a workspace, and when pulling it in via git it was treating it as one (and that was causing issues), but even after converting opte to a partial workspace (where ddi, opte-core, and opteadm are packages in the workspace), I still hit this linker error (versus when pulling opteadm in directly via a path).
I used truss -f -a -t exec to compare the exact invocations of ld for both the git case and path case. They are identical. But for some reason in the git case it wants nodev in the final object final, and I have no idea why. If I look at the binary for the path case it has nothing for the symbol nodev.
rpz@midgar:~/test-opte-workspace-path$ nm target/debug/deps/test_opte_workspace_path-7170a9230c9145c8 | grep nodev
rpz@midgar:~/test-opte-workspace-path$ Versus the git case:
rpz@midgar:~/test-opte-workspace$ nm target/debug/deps/test_opte_workspace-add43b47b53ce5cf | grep nodev
[28241] | 0| 0|NOTY |GLOB |0 |UNDEF |nodev
[16473] | 10033104| 40|FUNC |GLOB |0 |14 |nodev_getinfo
[17301] | 10033280| 47|FUNC |GLOB |0 |14 |nodev_ioctl
[21982] | 10033184| 35|FUNC |GLOB |0 |14 |nodev_power
[21108] | 10033232| 37|FUNC |GLOB |0 |14 |nodev_read
[14179] | 10033152| 32|FUNC |GLOB |0 |14 |nodev_reset
[15068] | 10033328| 37|FUNC |GLOB |0 |14 |nodev_writeThe only difference I can see between the two is that in the good (path) case it has incremental objects for ddi:
rpz@midgar:~/test-opte-workspace-path$ for f in $(find target -name '*.o'); do if nm $f | grep -w nodev; then echo "match: $f"; fi; done;
[71] | 0| 0|NOTY |GLOB |0 |UNDEF |nodev
match: target/debug/incremental/illumos_ddi_dki-gdb63ufcihta/s-g8kyytoc6h-1o0xg0m-2b4j88ufq9kop/zewzfia6tff8ao2.oBut there is no such thing for the bad (git) case.
This all feels like some inside baseball between cargo/rustc and how it builds up these objects and invokes the linker that is beyond my understanding on how to debug.
The real solution to this problem, once again, is that an API layer needs to be defined that insulates a client from all this crap. There is absolutely no reason for nodev() to be on the scene as that is a purely kernel driver concern. However, there is a hack that can be done in order to make forward progress until that separation happens.
rpz@midgar:~/test-opte-workspace$ cat .cargo/config
[build]
rustflags = ["-C", "link-args=-Wl,-znodefs"]This tells ld(1) to ignore the case where a symbol is undefined.
bnaecker
When using
opteadmas a Git dependency in a project's Cargo.toml, we get linking errors. Things seem to work fine if the dependency is specified as a local, path-dependency. To reproduce, create a new binary:With a
Cargo.tomlthat looks like this:I have the following minimal binary, in `src/main.rs:
Running the package's binary fails at link:
Cloning this repo and specifying it as a path dependency builds just fine. Changing the one dependency in
Cargo.tomlto:we can build and run the binary successfully: