[release-1.11] malloc: use jl_get_current_task to fix null check

[benlorenz]

Since 1.11.5 we are getting crashes from OpenMP threads of an external library which uses GMP numbers. This did not happen on 1.11.4. Version 1.12 and nightly are also unaffected.

Attaching gdb shows that the task struct is an invalid pointer (but not NULL):

(gdb) p ct
$1 = (jl_task_t *) 0xffffffffffffff90

This value seems to come from the jl_current_task macro

#define jl_current_task (container_of(jl_get_pgcstack(), jl_task_t, gcstack))

which doesn't check whether jl_get_pgcstack() returns NULL for foreign threads. Two lines further down is a check for ct != NULL but this cannot happen since container_of will just subtract a fixed offset from a NULL-pointer.

julia/src/gc.c

Lines 4122 to 4124 in 2d89891

	jl_task_t *ct = jl_current_task;
	void *data = malloc(sz);
	if (data != NULL && ct != NULL && ct->world_age) {

I changed the relevant two lines to use jl_get_current_task() which will properly return NULL if jl_get_pgcstack() returns NULL. This also aligns it with the current master:

julia/src/gc-stock.c

Lines 3741 to 3742 in 16eca6e

	jl_task_t *ct = jl_get_current_task();
	if (data != NULL && ct != NULL) {

This was probably introduced during the backport in #57880, @gbaraldi.

Note that this PR targets release-1.11, not sure if this is correct but there is no backports-release-1.11 branch right now.

cc: @lgoettgens
x-ref: oscar-system/Oscar.jl#4806

[lgoettgens]

Note that this PR targets release-1.11, not sure if this is correct but there is no backports-release-1.11 branch right now.

Maybe @KristofferC can comment on this

[gbaraldi]

LGTM branch target aside

[fingolfin]

There doesn't seem to be a backports 1.11.6 branch at this point. I realize 1.12 is coming closer, but it would be really unfortunate to leave 1.11.5 as last 1.11.x release with such a severe regression.

[KristofferC]

I'll make one soon