Weirdness (bug) in kernel function `jl_new_task`

[fingolfin]

In src/task.c, function jl_new_task does the following:

...
    if (ssize == 0) {
        ...
    }
    else {
        // user requested dedicated stack of a certain size
        if (ssize < MINSTKSZ)
            ssize = MINSTKSZ;
        t->bufsz = ssize;
        t->stkbuf = jl_alloc_fiber(&t->ctx, &t->bufsz, t);
        if (t->stkbuf == NULL)
            jl_throw(jl_memory_exception);
    }
...
    t->stkbuf = NULL;

So if ssize != 0 then t->stkbuf is allocated and care is taken to ensure it's non-zero, but then later it gets zeroed out unconditionally (no return or conditional statement statement comes in between, so I really think this is what happens).

This looks like a bug, presumably t->stkbuf. Even if it is intentional, then I'd think the first use of t->stkbuf inside the else {...} block should be rather replaced by a local variable, and a comment should be added why this makes sense.

[sgaure]

The pointer is stored in the context structure by jl_alloc_fiber(), so a local variable could have been used. It is merely slightly confusing, perhaps a leftover from an earlier version.

[fingolfin]

@sgaure Hmm, OK, but then something still doesn't up for me, perhaps you can clear it up for me: jl_alloc_fiber unconditionally allocates a new stack, right? Now these are all places I could find that set stkbuf:

1. in save_stack

    if (lastt->bufsz < nb) {
        buf = (void*)jl_gc_alloc_buf(ptls, nb);
        lastt->stkbuf = buf;
        lastt->bufsz = nb;
    }

2. in jl_finish_task

    if (t->copy_stack) // early free of stkbuf
        t->stkbuf = NULL;

3. in ctx_switch:

        if (t->stkbuf == NULL) {
            t->stkbuf = jl_alloc_fiber(&t->ctx, &t->bufsz, t);

4. in jl_new_task the two I listed at the top of this issue

5. in jl_init_root_task:

    if (always_copy_stacks) {
        ptls->current_task->copy_stack = 1;
        ptls->current_task->stkbuf = NULL;
        ptls->current_task->bufsz = 0;
    }
    else {
        ptls->current_task->copy_stack = 0;
        ptls->current_task->stkbuf = stack;
        ptls->current_task->bufsz = ssize;
    }

6. in jl_release_task_stack and sweep_stack_pools inside src/gc-stacks.c there is code which may reset it to NULL

Looking at these, it seems that task->stkbuf will never point at the stack allocated in that particular instance (except perhaps for root tasks, via the code in jl_init_root_task listed under 5, for which I did not analyze the callers). So, wouldn't ctx_switch then allocate another stack, because t->stkbuf would still be NULL?

What am I missing?

While at it, I am also wondering whether it is OK that save_stack uses jl_gc_alloc_buf instead of jl_malloc_stack (which is what the various jl_alloc_fiber implementations all seem to use)? (Clearly I don't understand this code, so my apologies if this is a really stupid question. It just stood out to me while collecting the snippets listed above)