Re: [patch 00/38] x86/retbleed: Call depth tracking mitigation

[Linus Torvalds]

On Mon, Jul 18, 2022 at 3:59 PM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
>
> That's bad. The function entry should be 16 byte aligned and as I just
> learned for AMD the ideal alignment would be possibly 32 byte as that's
> their I-fetch width.

In general, the L1 cache line size is likely the only "ideal" alignment.

Even if (I think) intel fetches just 16 bytes per cycle from the L1 I$
when decoding, being cacheline aligned still means that the L2->L1
transfer for the beginning of the function starts out better.

But Intel's current optimization many actually end sup having special rules:

   When executing code from the Decoded Icache, direct branches that
are mostly taken should have all their instruction bytes in a 64B
cache line and nearer the end of that cache line. Their targets should
be at or near the beginning of a 64B cache line.

   When executing code from the legacy decode pipeline, direct
branches that are mostly taken should have all their instruction bytes
in a 16B aligned chunk of memory and nearer the end of that 16B
aligned chunk. Their targets should be at or near the beginning of a
16B aligned chunk of memory.

So note how the branch itself should be at the end of the chunk, but
the branch _target_ should be at the beginning of the chunk. And the
chunk size is 16 bytes for decoding new instructions, and 64 bytes for
predecoded.

I suspect that for the kernel, and for the beginning of the function
(ie the target case), the 16-byte thing is still the main thing.
Because the L0 I$ ("uop cache", "Decoded Icache", whatever you want to
call it) is probably too small for a lot of kernel loads where user
space has flushed things in between system calls or faults.

Older versions of the intel code just said "All branch targets should
be 16-byte aligned".

So I think 16 bytes for function alignment ends up being what we
generally want, but yes, it's slowly changing. AMD fetches 32-byte
chunks, and Intel has that 64-bit predecode thing.

           Linus