I’ve just started toying around with Fuse. The possibilities of writing comprehensive filesystems in user space interests me, not the least because it makes development much easier before you move to kernel space.
I’ll say it right out, Fuse is certainly not the worst conceived / designed/ implemented piece of software by a long shot. The documentation isn’t great, but it’s not totally lacking either (which immediately puts Fuse a cut above a whole swathe of other open source projects). It’s not perfect however.
My main annoyance is the multi-threading support which, frankly, shouldn’t exist; i.e., it should be handled by the application and not the library. It is, generally, better when software layers are thin; it should be up to the app to do appropriate locking etc., and the library (in this case Fuse) should just specify in the documentation where the locking is needed (or rather, more usefully, where it is not needed). Ok, quibbles, and providing more functionality is hardly a crime; after all, the low-level layer is still available and (mostly) unobfuscated by the higher-level functionality.
The real issue is that linking against libfuse pulls in the pthread shared library as well, which is not so good. The pthread library incurs overhead (other than taking up memory/address space) because when it is loaded, locking and unlocking mutexes suddenly become real operations – which means fputc() and a whole bunch of other library functions become a weeny bit slower. (Don’t believe me? do “nm” on /lib/libc.so and /lib/libpthread.so and notice that they both define pthread_mutex_lock). Fuse should really be split into two librarys, libfuse and libfuse_mt, and only the latter should require pthreads. (Or, as I suggested earlier, just ditch the multi-thread functionality altogether, leave it up to the application to pull in pthreads if it really needs it).
Another annoyance is the fuse_chan_receive function:
int fuse_chan_recv(struct fuse_chan **ch, char *buf, size_t size);
Notice that first argument? Not just a pointer, it’s a pointer-to-a-pointer. Why, great Zeus, why??! The function is meant to receive data from a channel, in what crazy circumstance would you want it to change the channel pointer itself to point at something else? And in fact, the function delegates to a another function via a pointer (embedded in **ch) to a fuse_chan_ops structure; in the library itself, as far as I can see, there are only two implementations (one to receive directly from the kernel, and another called mt_chan_receive) and neither of them does actually alter the value of *ch. If you have a look at the implementation of fuse_loop() though you’ll notice the author clearly had the possibility in mind, since he copies the channel into a temporary pointer variable (tmpch) before using fuse_chan_recv – so yes, the bad API design has caused increased code complexity.
The comments above the declaration of fuse_chan_recv don’t give any clue as to why it would ever modify the *ch value and incidentally ch is incorrectly documented as being a “pointer to the channel” when it is of course really a “pointer to a pointer to the channel”.
Oh, and the mt_chan_recieve function I mentioned earlier – well, it presumably has something to do with the multi-threading support, but it’s hard to be sure because the function lacks anything bearing the slightest resemblance to a comment, and there certainly isn’t any other explanation of what it does nor how it works.