Re: [PATCH] ia64: enable HAVE_COPY_THREAD_TLS, switch to kernel_clone_args

[Christian Brauner]

On Thu, May 14, 2020 at 12:33:00PM +0200, Christian Brauner wrote:
> On Thu, May 14, 2020 at 12:21:13PM +0200, John Paul Adrian Glaubitz wrote:
> > On 5/14/20 12:19 PM, Christian Brauner wrote:
> > > Scratch that. It's even worse. On ia64 it is _invalid_ to pass a NULL
> > > stack. That's at least what the glibc assembly assumes:
> > > 
> > > 	cmp.eq p6,p0=0,in0
> > > 	cmp.eq p7,p0=0,in1
> > > 	mov r8=EINVAL
> > > 	mov out0=in3		/* Flags are first syscall argument.	*/
> > > 	mov out1=in1		/* Stack address.			*/
> > > (p6)	br.cond.spnt.many __syscall_error	/* no NULL function pointers */
> > > (p7)	br.cond.spnt.many __syscall_error	/* no NULL stack pointers */
> > > 	;;
> > > 	mov out2=in2		/* Stack size.				*/
> > > 
> > > so newer systemd just works by accident on ia64 if at all correctly
> > > afaict.
> > 
> > Hmm, interesting. I really wasn't aware of that. Thanks for the heads-up.
> > 
> > I'll ask Michael whether he can come up for a solution for that problem.
> > 
> > Maybe that's also why systemd crashes.
> 
> Do you have a very minimalistic ia64 userspace preferably without systemd where
> you could simply test. That should give us an idea whether things work:
> 
> #define _GNU_SOURCE
> #include <sys/wait.h>
> #include <sys/utsname.h>
> #include <sched.h>
> #include <string.h>
> #include <stdio.h>
> #include <stdlib.h>
> #include <unistd.h>
> #include <sys/mman.h>
> 
> #define STACK_SIZE (8 * 1024 * 1024) /* standard stack size for threads in glibc */
> 
> int main(int argc, char *argv[])
> {
> 	char *stack;
>         pid_t pid;
> 
> 	stack = mmap(NULL, STACK_SIZE, PROT_READ | PROT_WRITE,
> 		     MAP_PRIVATE | MAP_ANONYMOUS | MAP_STACK, -1, 0);
> 	if (stack == MAP_FAILED)
> 		exit(EXIT_FAILURE);
> 
>         /* 
> 	 * Note that legacy clone() has different argument ordering on
>          * different architectures so this won't work everywhere.
>          */
>         pid = syscall(189 /* __NR_clone2 */, SIGCHLD, stack, STACK_SIZE, NULL, NULL);

Please note that even on ia64 the stack grows down but in contrast to
all other architectures ia64 expects the _lowest_ address to be given
and will add STACK_SIZE to stack itself in copy_thread{_tls}(). (This is
all fixed in clone3() where you're always expected to pass down the
lowest address and the kernel figures it out for you.)

So this is intentional.

Christian