By user13334066 on Sht 04, 2006
It was a nice training, though i'm not completely reformed or anything. But it did leave me with a strong itch to write a threading library myself. It was an interesting experience indeed. So finally i ended up with a User Level thread library, implemented though SIGALRMs and the Solaris context routines. I have put it over here.
It offers the usual set of routines: thread creation and manipulation functions, and routines for mutexes, semaphores, condition variables, and barriers.
Having just the usual set seemed boring, so i have added a rather wierd feature: the user can register a callback function, which will be called during (just before, actually) each context switch. The function will be passed two pointer arguments by the thread library: pointer to the id of the outgoing thread, and to that of the incoming thread. So a user can actually manipulate and override the "stock" scheduling logic by modifying the incoming thread id.
Another use would be along, say, monitoring lines. Say there is memory corruption happening. Say under load. You don't know which thread is doing it. Maybe a third party plugin is responsible. In such cases, you can register a function which will keep monitoring the affected memory region. The moment corruption happens, it can print out the outgoing thread id - the culprit - and say do an abort.
The signal stack is an issue right now and should be fixed. Now there is only one signal stack shared across threads, and it is not a good idea at all. One thread can get scheduled out, then more interrupts come in, the signal stack gets all "altered", and when that earlier thread gets scheduled back in, it's gonna find a corrupted stack. Thinking of providing a separate (small) signal stack to each thread, on which the SIGALRMs will be handled. That way, a thread can be safely scheduled out while in "thread land", and safely scheduled back in, later.
I built/tested it on Solaris 2.6 as well as 8. Didn't bother with Solaris 10 because i would need to fix the makecontext() usage first. So right now it builds on 10, but crashes.
Next step is to fix some bugs and race conditions. And then the plan is to upgrade the whole thing by bringing in LWP support, and hence move to an M x N threading model.