cvs commit: src/sys/i386/i386 genassym.c machdep.c mp_machdep.c swtch.s sys_machdep.c src/sys/i386/include pcb.h pcpu.h segments.h sysarch.h

[Peter Wemm]

On Wed, Apr 13, 2005 at 10:57:17PM +0000, Peter Wemm wrote:
> peter       2005-04-13 22:57:17 UTC
> 
>   FreeBSD src repository
> 
>   Modified files:
>     sys/i386/i386        genassym.c machdep.c mp_machdep.c swtch.s 
>                          sys_machdep.c 
>     sys/i386/include     pcb.h pcpu.h segments.h sysarch.h 
>   Log:
>   Change the segment limits to 4GB, we set the user accessible bit on all
>   of the kernel address space already.  Intel recommend this anyway, because
>   using a non-4GB limit adds an additional clock cycle to address generation.
>   We were able to install 4GB segments into the LDT, so any limits we imposed
>   on %cs and %ds were academic anyway.  More importantly, this allows us to
>   make a page in the kernel readable to user applications, for holding things
>   like the signal trampoline and other fun things.
>   
>   Move the user %cs/%ds segments from the LDT to the GDT.  There was no good
>   reason for them to be there anyway.  The old LDT entries are still there
>   but we can now relax the restriction that prevented users from emptying
>   the default LDT entries.
>   
>   Putting user and kernel %cs and %ds together allows us to access the fast
>   sysenter/sysexit/syscall/sysret instructions.  syscall/sysret in particular
>   require that the user/kernel segments be laid out this way.  Reserve a slot
>   specifically for NDIS while here.
>   
>   Create two user controllable slots in the GDT that are context switched
>   with the (kernel) thread.  This allows user applications to set two
>   user privilige selectors to arbitary values.  Create
>   i386_set_fsbase(void *base) and friends. (get/set, fs/gs).  For i386,
>   %gs is used by tls and the thread libraries and this means that user
>   processes no longer have to have the cost of having a custom LDT, and
>   we will no longer to do a ldt switch when activating a kthread/ithread in
>   the usual case any more.
>   
>   In other words, we can now set the base address for %fs and %gs to arbitary
>   addresses without the pain of messing with ldt segments.

Please note that I have not tested this with SMP yet, but I am 99.9% sure
that it will work.

I have a bunch of patches for userland that will use this stuff.  It does
fall back to the old ldt method if you boot a kernel.old, so it should be
safe to commit these shortly.  I have patches for ld.so, libc and libpthread.
I have not tried libthr yet.

-Peter