TLS - implementing linux one in fbsd

[Divacky Roman]

> > Linux uses strict 1:1 threading so every thread is in fact process, so
> > thread creation is done using plain clone()/fork(). FreeBSD uses M:N
> > (including 1:1) threading. Threads are created via pthread_create() call to
> > threading library. In kernel there's thr_new() syscall or thread_create()
> > syscall. I didnt find the connection between threading library and kernel
> > but I assume its using one of the syscalls
> >
> The M:N and 1:1 threading in FreeBSD use different mechanisms to
> implement TLS, M:N implements it in userland, while 1:1 implements it in 
> kernel. the thr_new or thr_create are used for 1:1 threading, right 
> now libthr uses thr_new to atomically setup a thread, this includes,
> storing TID, setting TLS, and maybe signal mask( not implemented ) ,
> cpu affinity mask etcs(not implemented), scheduling scope, in one word,
> it is intended to map most part of pthread_attr into kernel world.
 
but on the kernel level the implementation must be the same.. I mean the
mangling of %gs. right?

 
> > For setting up the GDT for the thread Linux uses syscall set_thread_area()
> > (TODO - how exactly? its unclear what it does). I dont know how FreeBSD
> > does it but I think it might be done via params to the syscalls (TODO - how
> > is it done?)
> >
> If you use thr_new, it is not necessary to use set_thread_area, I am not sure
> you need to change TLS pointer again after the thread is created, I think
> only main thread may need this feature, in FreeBSD, setting thread's TLS
> pointer is via libc function: _set_tp(void *tp).

well.. in linux the thread creation and setting up the tls is done using
separate syscalls. I plan to extend clone() syscall to use thr_create() or
thr_new() (if the flags tell me its thread) but I am afraid I'l have to modify
those syscalls to not to setup TLS (some flag) because linux wants to set it
separately.

> > set/get tid - does it relate to TLS at all? I dont think so but you never
> > know. The tid thing is unclear to me. The clone() syscall is passed
> > CLONE_CHILD_SETTID & CLONE_CHILD_CLEARTID which should be mutually
> > exclusive. I dont believe much its a mistake.. but the code is clear:
> >  p->set_child_tid = (clone_flags & CLONE_CHILD_SETTID) ? child_tidptr :
> > NULL; p->clear_child_tid = (clone_flags & CLONE_CHILD_CLEARTID) ?
> > child_tidptr: NULL;
> >
> > kostik belousov pointed out that this is used for futexes, so not
> > interesting for this
> >
> >
> I think it is used for futex, and the childtid is use to implement
> pthread_join and garbage collection in thread library, the parent tid
> pointer (if I recall correctly) is used by parent thread to retrieve
> child tid.
 
this is the next step... I think all the magic is done in their libc (or
somewhere) and I basically just need to malloc some space for this info
and clear/set it on proces creation/exit
 
> > Possible mapping from Linux to FreeBSD:
> >
> > To me it seems that the the set_thread_area() syscall is used in the
> > process of thread creation to set where the tls is stored. In FreeBSD we
> > use cpu_set_user_tls() for this. So it might be enough to just wrap call to
> > cpu_set_user_tls() into the syscall.
> cpu_set_user_tls is used by thr_new syscall internally to setup TLS pointer
> before executing user code, the thr_new syscall's only user is libthr.

the cpu_set_user_tls() is then what I need I think... maybe some modifications
needed but it shuold be basiscally the thing.

the linux syscall set_thread_area() just loads GDT with that info.. thats the
same like ours cpu_set_use_tls(), right?

thnx for your information!

roman