Re: [PATCH] x86/paravirt: Add missing noinstr to arch_local*() helpers

[Marco Elver]

On Wed, 5 Aug 2020 at 16:36, Marco Elver <elver@xxxxxxxxxx> wrote:
>
> On Wed, 5 Aug 2020 at 16:17, <peterz@xxxxxxxxxxxxx> wrote:
> >
> > On Wed, Aug 05, 2020 at 04:12:37PM +0200, peterz@xxxxxxxxxxxxx wrote:
> > > On Wed, Aug 05, 2020 at 03:59:40PM +0200, Marco Elver wrote:
> > > > On Wed, Aug 05, 2020 at 03:42PM +0200, peterz@xxxxxxxxxxxxx wrote:
> > >
> > > > > Shouldn't we __always_inline those? They're going to be really small.
> > > >
> > > > I can send a v2, and you can choose. For reference, though:
> > > >
> > > >     ffffffff86271ee0 <arch_local_save_flags>:
> > > >     ffffffff86271ee0:       0f 1f 44 00 00          nopl   0x0(%rax,%rax,1)
> > > >     ffffffff86271ee5:       48 83 3d 43 87 e4 01    cmpq   $0x0,0x1e48743(%rip)        # ffffffff880ba630 <pv_ops+0x120>
> > > >     ffffffff86271eec:       00
> > > >     ffffffff86271eed:       74 0d                   je     ffffffff86271efc <arch_local_save_flags+0x1c>
> > > >     ffffffff86271eef:       0f 1f 44 00 00          nopl   0x0(%rax,%rax,1)
> > > >     ffffffff86271ef4:       ff 14 25 30 a6 0b 88    callq  *0xffffffff880ba630
> > > >     ffffffff86271efb:       c3                      retq
> > > >     ffffffff86271efc:       0f 1f 44 00 00          nopl   0x0(%rax,%rax,1)
> > > >     ffffffff86271f01:       0f 0b                   ud2
> > >
> > > >     ffffffff86271a90 <arch_local_irq_restore>:
> > > >     ffffffff86271a90:       53                      push   %rbx
> > > >     ffffffff86271a91:       48 89 fb                mov    %rdi,%rbx
> > > >     ffffffff86271a94:       0f 1f 44 00 00          nopl   0x0(%rax,%rax,1)
> > > >     ffffffff86271a99:       48 83 3d 97 8b e4 01    cmpq   $0x0,0x1e48b97(%rip)        # ffffffff880ba638 <pv_ops+0x128>
> > > >     ffffffff86271aa0:       00
> > > >     ffffffff86271aa1:       74 11                   je     ffffffff86271ab4 <arch_local_irq_restore+0x24>
> > > >     ffffffff86271aa3:       0f 1f 44 00 00          nopl   0x0(%rax,%rax,1)
> > > >     ffffffff86271aa8:       48 89 df                mov    %rbx,%rdi
> > > >     ffffffff86271aab:       ff 14 25 38 a6 0b 88    callq  *0xffffffff880ba638
> > > >     ffffffff86271ab2:       5b                      pop    %rbx
> > > >     ffffffff86271ab3:       c3                      retq
> > > >     ffffffff86271ab4:       0f 1f 44 00 00          nopl   0x0(%rax,%rax,1)
> > > >     ffffffff86271ab9:       0f 0b                   ud2
> > >
> > >
> > > Blergh, that's abysmall. In part I suspect because you have
> > > CONFIG_PARAVIRT_DEBUG, let me try and untangle that PV macro maze.
> >
> > Yeah, look here:
> >
> > 0000 0000000000462149 <arch_local_save_flags>:
> > 0000   462149:  ff 14 25 00 00 00 00    callq  *0x0
> > 0003                    46214c: R_X86_64_32S    pv_ops+0x120
> > 0007   462150:  c3                      retq
> >
> >
> > That's exactly what I was expecting.
>
> Ah, for some reason the __always_inline version does *not* work with
> KCSAN -- I'm getting various warnings, including the same lockdep
> warning. I think there is some weirdness when this stuff gets inlined
> into instrumented functions. At least with KCSAN, when any accesses
> here are instrumented, and then KCSAN disable/enables interrupts,
> things break. So, these functions should never be instrumented,
> noinstr or not. Marking them 'inline noinstr' seems like the safest
> option. Without CONFIG_PARAVIRT_DEBUG, any compiler should hopefully
> inline them?

Oh well, it seems that KCSAN on syzbot still crashes even with this
"fix". It's harder to reproduce though, and I don't have a clear
reproducer other than "fuzz the kernel" right now. I think the new IRQ
state tracking code is still not compatible with KCSAN, even though we
thought it would be. Most likely there are still ways to get recursion
lockdep->KCSAN. An alternative would be to deal with the recursion
like we did before, instead of trying to squash all of it. I'll try to
investigate -- Peter, if you have ideas, help is appreciated.

Thanks,
-- Marco