Re: [PATCH] x86/entry/64: Remove duplicate syscall table for fast path
From: Andy Lutomirski
Date: Wed Dec 09 2015 - 16:16:19 EST
On Wed, Dec 9, 2015 at 1:08 PM, Brian Gerst <brgerst@xxxxxxxxx> wrote:
> On Wed, Dec 9, 2015 at 1:53 PM, Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:
>> On Wed, Dec 9, 2015 at 5:02 AM, Brian Gerst <brgerst@xxxxxxxxx> wrote:
>>> Instead of using a duplicate syscall table for the fast path, create stubs for
>>> the syscalls that need pt_regs that save the extra registers if a flag for the
>>> slow path is not set.
>>> Signed-off-by: Brian Gerst <brgerst@xxxxxxxxx>
>>> To: Andy Lutomirski <luto@xxxxxxxxxxxxxx>
>>> Cc: Andy Lutomirski <luto@xxxxxxxxxx>
>>> Cc: the arch/x86 maintainers <x86@xxxxxxxxxx>
>>> Cc: Linux Kernel Mailing List <linux-kernel@xxxxxxxxxxxxxxx>
>>> Cc: Borislav Petkov <bp@xxxxxxxxx>
>>> Cc: FrÃdÃric Weisbecker <fweisbec@xxxxxxxxx>
>>> Cc: Denys Vlasenko <dvlasenk@xxxxxxxxxx>
>>> Cc: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
>>> Applies on top of Andy's syscall cleanup series.
>> A couple questions:
>>> @@ -306,15 +306,37 @@ END(entry_SYSCALL_64)
>>> - * Syscalls marked as needing ptregs that go through the fast path
>>> - * land here. We transfer to the slow path.
>>> + * Syscalls marked as needing ptregs land here.
>>> + * If we are on the fast path, we need to save the extra regs.
>>> + * If we are on the slow path, the extra regs are already saved.
>>> - DISABLE_INTERRUPTS(CLBR_NONE)
>>> - TRACE_IRQS_OFF
>>> - addq $8, %rsp
>>> - jmp entry_SYSCALL64_slow_path
>>> + movq PER_CPU_VAR(cpu_current_top_of_stack), %r10
>>> + testl $TS_SLOWPATH, ASM_THREAD_INFO(TI_status, %r10, 0)
>>> + jnz 1f
>> OK (but see below), but why not do:
>> addq $8, %rsp
>> jmp entry_SYSCALL64_slow_path
> I've always been adverse to doing things like that because it breaks
> call/return branch prediction.
I'd agree with you there except that the syscalls in question really
don't matter for performance enough that we should worry about a
handful of cycles from a return misprediction. We're still avoiding
IRET regardless (to the extent possible), and that was always the
> Also, are there any side effects to calling enter_from_user_mode()
> more than once?
A warning that invariants are broken if you have an appropriately
>> here instead of the stack munging below?
>>> + subq $SIZEOF_PTREGS, %r10
>>> + SAVE_EXTRA_REGS base=r10
>>> + movq %r10, %rbx
>>> + call *%rax
>>> + movq %rbx, %r10
>>> + RESTORE_EXTRA_REGS base=r10
>>> + ret
>>> + jmp *%rax
> After some thought, that can be simplified. It's only executed on the
> fast path, so pt_regs is at 8(%rsp).
>> Also, can we not get away with keying off rip or rsp instead of
>> ti->status? That should be faster and less magical IMO.
> Checking if the return address is the instruction after the fast path
> dispatch would work.
> Simplified version:
> cmpl $fast_path_return, (%rsp)
Does that instruction actually work the way you want it to? (Does it
link?) I think you might need to use leaq the way I did in my patch.
> jne 1f
> SAVE_EXTRA_REGS offset=8
> call *%rax
> RESTORE_EXTRA_REGS offset=8
> jmp *%rax
This'll work, I think, but I still think I prefer keeping as much
complexity as possible in the slow path. I could be convinced
otherwise, though -- this variant is reasonably clean.
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/