are all the changes to liblinear coming from upstream?

No, they're local modifications. There are liblinear forks on github, but I didn't find a convenient place to make changes, and actually wasn't 100% sure of the authoritative source.

Liblinear has recently been patched to handle sample weights #5274. Yet it was smaller than this patch.

To better understand what this patch does, it:

1. Adds wrapper functions, to start/join threads using pthreads, that support Windows, OS X and other Unixes, as well as wrappers for mutexes and semaphores.
2. Targets only two liblinear solvers and cases where there are more than two classes.
3. Saves memory by performing a transposition that only applies to these two solvers once instead of once per invocation of train_one.
4. Starts up n threads, and has them chew away at the independent one-vs-rest training for each of the classes until they're all done.

Since there already is an n_jobs parameter on the LogisticRegression class, and to give the API control over whether threads are used, and how many, I introduced a new parameter on the LR class called n_threads, since this parallelization may or may not be desirable in conjunction with the grid search parallelization provided by n_jobs. If other scikit-learn classes than LogisticRegression use liblinear, presumably the n_threads parameter is defaulting to 1, because it compiles fine.

It's been tested many times on OS X and Ubuntu 15.10 on datasets that used to take over 7 days to train, and which now takes about 11 hours using 48 cores.

I added a compile switch to turn off the pthreads dependency completely, in case there are compilation issues under some circumstances.

There is also parallelization available for single runs of train_one, e.g. if one only has 2 target classes, but these are still somewhat experimental. The two best contenders in that space that I've found are Shotgun and Bundle CDN. But in my case, I'm happy with just parallelizing across different classes.

Since as our docstring now states, liblinear is only good for small datasets, I don't think we need to add this as other solvers handle parallelism.