mirror of
https://github.com/llvm/llvm-project.git
synced 2025-05-11 12:56:08 +00:00
8cc49bec2e
This fixes an flakyness is all gdb-remote tests. These tests have been (mildly) flaky since we started using "localhost" instead of 127.0.0.1 in the test suite. The reason is that lldb-server needs to create two sockets (v4 and v6) to listen for localhost connections. The algorithm it uses first tries to select a random port (bind(localhost:0)) for the first address, and then bind the same port for the second one. The creating of the second socket can fail as there's no guarantee that port will be available -- it seems that the (linux) kernel tries to choose an unused port for the first socket (I've had to create thousands of sockets to reproduce this reliably), but this can apparantly fail when the system is under load (and our test suite creates a _lot_ of sockets). The socket creationg operation is considered successful if it creates at least one socket is created, but the test harness has no way of knowing which one it is, so it can end up connecting to the wrong address. I'm not aware of a way to atomically create two sockets bound to the same port. One way to fix this would be to make lldb-server report the address is it listening on instead of just the port. However, this would be a breaking change and it's not clear to me that's worth it (the algorithm works pretty well under normal circumstances). Instead, this patch sidesteps that problem by using "reverse" connections. This way, the test harness is responsible for creating the listening socket so it can pass the address that it has managed to open. It also results in much simpler code overall. To preserve test coverage for the named pipe method, I've moved the relevant code to a dedicated test. To avoid original problem, this test passes raw addresses (as obtained by getaddrinfo(localhost)) instead of "localhost". Differential Revision: https://reviews.llvm.org/D90313