On Mon, Aug 05, 2019 at 10:12:07AM +0000, Pascal Van Leeuwen wrote:
> > From: Antoine Tenart <antoine.tenart@xxxxxxxxxxx>
> > On Mon, Aug 05, 2019 at 09:48:13AM +0000, Pascal Van Leeuwen wrote:
> > > > On Wed, Jul 31, 2019 at 05:29:19PM +0200, Pascal van Leeuwen wrote:
> > > > >
> > > > > - /* Release engine from reset */
> > > > > - val = readl(EIP197_PE(priv) + ctrl);
> > > > > - val &= ~EIP197_PE_ICE_x_CTRL_SW_RESET;
> > > > > - writel(val, EIP197_PE(priv) + ctrl);
> > > > > + for (pe = 0; pe < priv->config.pes; pe++) {
> > > > > + base = EIP197_PE_ICE_SCRATCH_RAM(pe);
> > > > > + pollcnt = EIP197_FW_START_POLLCNT;
> > > > > + while (pollcnt &&
> > > > > + (readl_relaxed(EIP197_PE(priv) + base +
> > > > > + pollofs) != 1)) {
> > > > > + pollcnt--;
> > > >
> > > > You might want to use readl_relaxed_poll_timeout() here, instead of a
> > > > busy polling.
> > > >
> > > Didn't know such a thing existed, but I also wonder how appropriate it
> > > is in this case, condering it measures in whole microseconds, while the
> > > response time I'm expecting here is in the order of a few dozen nano-
> > > seconds internally ... i.e. 1 microsecond is already a *huge* overkill.
> > >
> > > The current implementation runs that loop for only 16 iterations which
> > > should be both more than sufficient (it probably could be reduced
> > > further, I picked 16 rather arbitrarily) and at the same time take so
> > > few cycles on the CPU that I doubt it is worthwhile to reschedule/
> > > preempt/whatever?
> >
> > Your choice, I was just making a suggestion :)
> >
> After reading the implementation code, I think using it with a
> sleep_us value of 0 and a time-out value of 1 may be reasonable here,
> as time-out means some serious error that shouldn't normally happen.
> I just wouldn't want any microsecond delays for the "normal" case.
Using this function that is designed to sleep with a delay of 0 and
designed to timeout with a value of 1 does not seem to follow what the
function is designed for :) It could work but I would suggest to keep
the polling as it is in the patch in this case.
Thanks,
Antoine
--
Antoine Ténart, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com
