In cgo mode for UNIX platforms, we compute the g0 stack bounds by querying the stack size from pthread_get_stacksize, subtracting that from the current SP in mstart, and adding a 1k buffer. However, on my Linux laptop, over 4k of the stack has already been consumed by the time we reach mstart, so the runtime sets the bottom of the g0 stack to be below the mapped memory. As a result, if we actually overflow the g0 stack, we get a segfault instead of a useful message with a traceback.
We should use pthread_get_stack instead to get both stack bounds.
Discovered when writing a test of overflowing the g0 stack for #21382, which I had assumed would pass on linux/amd64 but didn't.
In cgo mode for UNIX platforms, we compute the g0 stack bounds by querying the stack size from
pthread_get_stacksize, subtracting that from the current SP inmstart, and adding a 1k buffer. However, on my Linux laptop, over 4k of the stack has already been consumed by the time we reachmstart, so the runtime sets the bottom of the g0 stack to be below the mapped memory. As a result, if we actually overflow the g0 stack, we get a segfault instead of a useful message with a traceback.We should use
pthread_get_stackinstead to get both stack bounds.Discovered when writing a test of overflowing the g0 stack for #21382, which I had assumed would pass on linux/amd64 but didn't.