Creating a task with a custom stack size fails with `julia-debug`

[vchuravy]

On 1.10

function with_stack(f, n, pool=:default)
           @assert n > 0
           t = Core.Task(f, n)
           t.sticky=false
           Base.Threads._spawn_set_thrpool(t, pool)
           schedule(t)
           return fetch(t)
end

with_stack(()->nothing, 16_000_000)

[410693] signal (11.1): Segmentation fault
in expression starting at REPL[2]:1
fish: Job 1, '~/src/julia-1.10/julia' terminated by signal SIGSEGV (Address boundary error)

Make.user:

~/s/julia-1.10 (release-1.10)> cat Make.user
LLVM_ASSERTIONS=1
FORCE_ASSERTIONS=1
override CC=gcc-14
override CXX=g++-14

# default to a debug build for better line number reporting
override JULIA_BUILD_MODE=debug

WITH_TRACY=1

Thread 1 received signal SIGSEGV, Segmentation fault.
0x00007fea4287887f in jl_start_fiber_set (t=<error reading variable: Cannot access memory at address 0xffffffffffffffd8>) at /home/vchuravy/src/julia-1.10/src/task.c:1464
1464	    asm volatile (
(rr) rsi
0x00007fea4287887c in jl_start_fiber_set (t=0x7fe9e3468a10) at /home/vchuravy/src/julia-1.10/src/task.c:1464

Inspection shows that:

1459	    char *stk = ((char**)&t->ctx)[0];
1460	    size_t ssize = ((size_t*)&t->ctx)[1];

are both null.

Setting a breakpoint leads to:

#1  0x00007fea4287821a in ijl_new_task (start=0x7fe9e6e45d30, completion_future=0x7fea2361b950, ssize=16000000) at /home/vchuravy/src/julia-1.10/src/task.c:1099
1099	        memset(&t->ctx, 0, sizeof(t->ctx));

julia/src/task.c

Line 1099 in 95f30e5

memset(&t->ctx, 0, sizeof(t->ctx));

[vchuravy]

@vtjnash things work if I just remove the memset(&t->ctx, 0, sizeof(t->ctx)); but I assume it is there for a good reason?

[gbaraldi]

Is it overwriting something? That code is quite old and probably not really tested

[vtjnash]

It shouldn't be there, since we allocated this part of the ctx much earlier now

[vchuravy]

The code got deleted for 1.12 and above in #55515)

[giordano]

I presume this was fixed by #59917 (wasn't closed automatically because the PR wasn't against the default branch)