Date:   Fri, 17 Jul 2020 14:56:55 -0700
From:   Andy Lutomirski <luto@...nel.org>
To:     Thomas Gleixner <tglx@...utronix.de>
Cc:     Kees Cook <keescook@...omium.org>,
        LKML <linux-kernel@...r.kernel.org>, X86 ML <x86@...nel.org>,
        linux-arch <linux-arch@...r.kernel.org>,
        Will Deacon <will@...nel.org>, Arnd Bergmann <arnd@...db.de>,
        Mark Rutland <mark.rutland@....com>,
        Keno Fischer <keno@...iacomputing.com>,
        Paolo Bonzini <pbonzini@...hat.com>,
        kvm list <kvm@...r.kernel.org>,
        Gabriel Krisman Bertazi <krisman@...labora.com>
Subject: Re: [patch V3 01/13] entry: Provide generic syscall entry functionality

On Fri, Jul 17, 2020 at 12:29 PM Thomas Gleixner <tglx@...utronix.de> wrote:
>
> Kees Cook <keescook@...omium.org> writes:
> > On Thu, Jul 16, 2020 at 11:55:59PM +0200, Thomas Gleixner wrote:
> >> Kees Cook <keescook@...omium.org> writes:
> >> >> +/*
> >> >> + * Define dummy _TIF work flags if not defined by the architecture or for
> >> >> + * disabled functionality.
> >> >> + */
> >> >
> >> > When I was thinking about this last week I was pondering having a split
> >> > between the arch-agnositc TIF flags and the arch-specific TIF flags, and
> >> > that each arch could have a single "there is agnostic work to be done"
> >> > TIF in their thread_info, and the agnostic flags could live in
> >> > task_struct or something. Anyway, I'll keep reading...
> >>
> >> That's going to be nasty. We rather go and expand the TIF storage to
> >> 64bit. And then do the following in a generic header:
> >
> > I though the point was to make the TIF_WORK check as fast as possible,
> > even on the 32-bit word systems. I mean it's not a huge performance hit,
> > but *shrug*
>
> For 64bit it's definitely faster to have 64bit flags.
>
> For 32bit it's debatable whether having to fiddle with two words and
> taking care about ordering is faster or not. It's a pain on 32bit in any
> case.
>
> But what we can do is distangle the flags into those which really need
> to be grouped into a single 32bit TIF word and architecture specific
> ones which can live in a different place.
>
> Looking at x86 which has the most pressure on TIF bits:
>
> Real TIF bits
> TIF_SYSCALL_TRACE       Entry/exit
> TIF_SYSCALL_TRACEPOINT  Entry/exit
> TIF_NOTIFY_RESUME       Entry/exit
> TIF_SIGPENDING          Entry/exit
> TIF_NEED_RESCHED        Entry/exit
> TIF_SINGLESTEP          Entry/exit
> TIF_SYSCALL_EMU         Entry/exit
> TIF_SYSCALL_AUDIT       Entry/exit
> TIF_SECCOMP             Entry/exit
> TIF_USER_RETURN_NOTIFY  Entry/exit      (x86 specific)
> TIF_UPROBE              Entry/exit
> TIF_PATCH_PENDING       Entry/exit
> TIF_POLLING_NRFLAG      Scheduler (related to TIF_NEED_RESCHED)
> TIF_MEMDIE              Historical, but not required to be a real one
> TIF_FSCHECK             Historical, but not required to be a real one
>
> Context switch group
> TIF_NOCPUID             X86 Context switch
> TIF_NOTSC               X86 Context switch
> TIF_SLD                 X86 Context switch
> TIF_BLOCKSTEP           X86 Context switch
> TIF_IO_BITMAP           X86 Context switch + Entry/exit (see below)
> TIF_NEED_FPU_LOAD       X86 Context switch + Entry/exit (see below)
> TIF_SSBD                X86 Context switch
> TIF_SPEC_IB             X86 Context switch
> TIF_SPEC_FORCE_UPDATE   X86 Context switch
>
> No group requirements
> TIF_IA32                X86 random
> TIF_FORCED_TF           X86 random
> TIF_LAZY_MMU_UPDATES    X86 random
> TIF_ADDR32              X86 random
> TIF_X32                 X86 random
>
> So the only interesting ones are TIF_IO_BITMAP and TIF_NEED_FPU_LOAD,
> but those are really not required to be in the real TIF group because
> they are independently evaluated _after_ the real TIF flags on exit to
> user space and that requires a reread of the flags anyway. So if we put
> the context switch and the random bits into a seperate word right after
> thread_info->flags then the second word is in the same cacheline and it
> wont matter. That way we gain plenty of free bits on 32 bit and have no
> dependency between the two words at all.
>
> The alternative is to play nasty games with TIF_IA32, TIF_ADDR32 and
> TIF_X32 to free up bits for 32bit and make the flags field 64 bit on 64
> bit kernels, but I prefer to do the above seperation.

I'm all for cleaning it up, but I don't think any nasty games would be
needed regardless.  IMO at least the following flags are nonsense and
don't belong in TIF_anything at all:

TIF_IA32, TIF_X32: can probably be deleted.  Someone would just need
to finish the work.
TIF_ADDR32: also probably removable, but I'm less confident.
TIF_FORCED_TF: This is purely a ptrace artifact and could easily go
somewhere else entirely.

So getting those five bits back would be straightforward.

FWIW, TIF_USER_RETURN_NOTIFY is a bit of an odd duck: it's an
entry/exit word *and* a context switch word.  The latter is because
it's logically a per-cpu flag, not a per-task flag, and the context
switch code moves it around so it's always set on the running task.
TIF_NEED_RESCHED is sort of in this category, too.  We could introduce
a percpu entry_exit_work field to simplify this at some small
performance cost.  TIF_POLLING_NRFLAG would go along with it.  (The
latter does not, strictly speaking, belong as a TIF_ flag at all, but
it does need to be in the same atomic word as TIF_NEED_RESCHED.)
Making this change would arguably be a decent cleanup.

--Andy

Powered by blists - more mailing lists