Re: [PATCH 09/18] soc: qcom: ipa: GSI transactions
From: Evan Green
Date: Mon May 20 2019 - 13:39:35 EST
On Mon, May 20, 2019 at 9:50 AM Alex Elder <elder@xxxxxxxxxx> wrote:
>
> On 5/20/19 11:34 AM, Evan Green wrote:
> > On Mon, May 20, 2019 at 7:44 AM Alex Elder <elder@xxxxxxxxxx> wrote:
> >>
> >> On 5/20/19 9:43 AM, Arnd Bergmann wrote:
> >>> I have no idea how two 8-bit assignments could do that,
> >>> it sounds like a serious gcc bug, unless you mean two
> >>> 8-byte assignments, which would be within the range
> >>> of expected behavior. If it's actually 8-bit stores, please
> >>> open a bug against gcc with a minimized test case.
> >>
> >> Sorry, it's 8 *byte* assignments, not 8 bit. -Alex
> >
> > Is it important to the hardware that you're writing all 128 bits of
>
> No, it is not important in the ways you are describing.
>
> We're just geeking out over how to get optimal performance.
> A single 128-bit write is nicer than two 64-bit writes,
> or more smaller writes.
>
> The hardware won't touch the TRE until the doorbell gets
> rung telling it that it is permitted to do so. The doorbell
> is an I/O write, which implies a memory barrier, so the entire
> TRE will be up-to-date in memory regardless of whether we
> write it 128 bits or 8 bits at a time.
>
Ah, understood. Carry on!