Messages in this thread | | | Subject | Re: [PATCH 03/15] x86: remove early_gdt_descr reference | From | James Bottomley <> | Date | Mon, 09 Jun 2008 12:40:02 -0500 |
| |
On Mon, 2008-06-09 at 14:23 -0300, Glauber Costa wrote: > James Bottomley wrote: > > On Mon, 2008-06-09 at 12:49 -0300, Glauber Costa wrote: > >> James Bottomley wrote: > >>> On Mon, 2008-06-09 at 11:16 -0300, Glauber Costa wrote: > >>>> since we use switch_to_new_gdt, there is no point > >>>> in assigning early_gdt_descr except for the first > >>>> assignment, which is done manually. > >>> What makes you think you can do this? If you don't update the early > >>> boot gdt, they all end up using the Boot CPU one. The problem with this > >>> is that there's a time from start_secondary to switch_to_new_gdt where > >>> the per cpu selector (%fs) and the pda selector (%gs) are those of the > >>> boot CPU. The former isn't a problem but the CPU number is in the > >>> latter, and it's used in that path before we get to the initialisation. > >> You are right, I missed it. > >> > >> However, it only seem to be used in cpu_init, and very early. Sure there > >> are some users _before_ we load the new gdt, but nothing prevents them > >> to be moved after it. (Of course, this patch is wrong anyway). > >> > >> And if we do that, we can even take the %fs loading out of head_32.S > >> Of course, it's only valid if those are indeed the only early users of it. > >> > >> Is there any other use I'm missing? > > > > Well, %fs loading there is done for the boot CPU. To eliminate that you > > have to not only verify that start_secondary doesn't use anything in > > per_cpu areas, but also verify that nothing in start_kernel() up until > > boot_cpu_init() does ... That's a lot of smp_processor_id() references > > to convert. > Yes, after a second look, it would be tricky indeed. But only for cpu0. > For all the others, I still think we could get rid of the problem by > switching to the new gdt earlier in cpu_init. > > What do you think?
Operating a CPU with a bogus GDT is very fragile. You can fix all the current issues with the secondary CPUs, but it gives a critical section within which none of the per_cpu operations will work. It only takes one patch violating this rule and we have a very subtle bug introduced.
It looks to me like the better fix might be just to initialise the gdt completely and properly in do_boot_cpu and just have the single switch in head_32.S be the correct one. That way there's no problem critical region.
James
| |