4 ms·

Alright, case study: I want to write a function 'F', that needs to initialise some state by calling 'I' the first time 'F' is called. 'I' itself may call 'F' ho

Alright, case study: I want to write a function 'F', that needs to initialise some state by calling 'I' the first time 'F' is called. 'I' itself may call 'F' however; and this needs to be handled specially in 'F' on the second call so that 'I' doesn't get called a second time. Until the first invocation of 'I' has returned (i.e., state has been initialised), no one else but the thread of the first call to 'F' should be allowed to call 'F'. Only the first call to 'F' must call 'I', not any other call. How do I solve this without using recursive mutex? This may sound like some nightmare architecture, but I ended up in this situation when writing a simple thread-safe malloc override using LD_PRELOAD.

LegionMammal978 4y ago
Assuming that 'I' can receive a payload, and that it doesn't need to call 'F' directly, you could write a function 'G' that takes a pointer to the locked state of 'F'. Then, 'F' can lock a non-recursive mutex and call 'G', 'G' can pass that pointer to 'I', and 'I' can pass the pointer back to 'G'. (If 'I' cannot receive a payload, you could also write a variant where 'G' simply assumes that the state is locked, and accesses it directly.) That's basically what this article is suggesting. Otherwise, the only thing I could think of is a thread-local boolean to tell 'F' that it is being called from 'I'. Which is, as I understand it, basically what a recursive mutex does, by remembering the thread that locked it. But it's still technically a different construct. The tricky part of the problem under this condition is that only the recursive call of 'F' from 'I' can proceed, so somehow that call must be distinguished from an identical call to 'F' from a different thread. Overall, things are much trickier, especially in multithreaded code, when you can't directly thread state through your functions. If there's one point I'll give in favor of the FP philosophy, it's that.
gpderetta 4y ago
Have both F and I call a private F_unlocked. I assume the issue is you can't (easily) change I (for example F is malloc). Then any solution in practice would look like a recursive mutex. That's alright, recursive mutexes are bad, but one legitimate use case is retrofitting thread safety on top on a non thread safe interface (infamously xlib was made thread safe by just bolting on a global recursive mutex).

LegionMammal978 4y ago

Assuming that 'I' can receive a payload, and that it doesn't need to call 'F' directly, you could write a function 'G' that takes a pointer to the locked state of 'F'. Then, 'F' can lock a non-recursive mutex and call 'G', 'G' can pass that pointer to 'I', and 'I' can pass the pointer back to 'G'. (If 'I' cannot receive a payload, you could also write a variant where 'G' simply assumes that the state is locked, and accesses it directly.) That's basically what this article is suggesting. Otherwise, the only thing I could think of is a thread-local boolean to tell 'F' that it is being called from 'I'. Which is, as I understand it, basically what a recursive mutex does, by remembering the thread that locked it. But it's still technically a different construct. The tricky part of the problem under this condition is that only the recursive call of 'F' from 'I' can proceed, so somehow that call must be distinguished from an identical call to 'F' from a different thread. Overall, things are much trickier, especially in multithreaded code, when you can't directly thread state through your functions. If there's one point I'll give in favor of the FP philosophy, it's that.

gpderetta 4y ago

Have both F and I call a private F_unlocked. I assume the issue is you can't (easily) change I (for example F is malloc). Then any solution in practice would look like a recursive mutex. That's alright, recursive mutexes are bad, but one legitimate use case is retrofitting thread safety on top on a non thread safe interface (infamously xlib was made thread safe by just bolting on a global recursive mutex).