Messages in this thread Patch in this message | | | From | Nikhil Gupta <> | Subject | [PATCH net-next] ptp_qoriq: fix latency in ptp_qoriq_adjtime() operation. | Date | Mon, 16 Jan 2023 09:28:33 +0000 |
| |
-----Original Message----- From: Jakub Kicinski <kuba@kernel.org> Sent: Thursday, January 12, 2023 10:39 AM To: Nikhil Gupta <nikhil.gupta@nxp.com> Cc: linux-arm-kernel@lists.infradead.org; Y.B. Lu <yangbo.lu@nxp.com>; netdev@vger.kernel.org; linux-kernel@vger.kernel.org; Vakul Garg <vakul.garg@nxp.com>; Rajan Gupta <rajan.gupta@nxp.com> Subject: [EXT] Re: [PATCH] ptp_qoriq: fix latency in ptp_qoriq_adjtime() operation.
Caution: EXT Email
please put [PATCH net-next] in the subject.
On Tue, 10 Jan 2023 17:00:24 +0530 nikhil.gupta@nxp.com wrote: > From: Nikhil Gupta <nikhil.gupta@nxp.com> > > 1588 driver loses about 1us in adjtime operation at PTP slave. > This is because adjtime operation uses a slow non-atomic > tmr_cnt_read() followed by tmr_cnt_write() operation.
So far so good..
> In the above sequence, since the timer counter operation loses about 1us.
s/operation/keeps incrementing after the read/ ?
but frankly I don't think this sentence adds much
> Instead, tmr_offset register should be programmed with the delta > nanoseconds
missing full stop at the end. You should describe what the tmr_offset register does. [Nikhil] : current time is calculated by adding TMROFF_H/L with the timer's counter TMR_CNT_H/L register.
> This won't lead to timer counter stopping and losing time while > tmr_cnt_write() is being done.
Stopping? The timer was actually stopping?
> This Patch adds api for tmr_offset_read/write to program the
Use imperative mood.
> delta nanoseconds in the Timer offset Register. > > Signed-off-by: Nikhil Gupta <nikhil.gupta@nxp.com> > Reviewed-by: Yangbo Lu <yangbo.lu@nxp.com> > --- > drivers/ptp/ptp_qoriq.c | 36 ++++++++++++++++++++++++++++++------ > 1 file changed, 30 insertions(+), 6 deletions(-) > > diff --git a/drivers/ptp/ptp_qoriq.c b/drivers/ptp/ptp_qoriq.c index > 08f4cf0ad9e3..5b6ea6d590be 100644 > --- a/drivers/ptp/ptp_qoriq.c > +++ b/drivers/ptp/ptp_qoriq.c > @@ -48,6 +48,29 @@ static void tmr_cnt_write(struct ptp_qoriq *ptp_qoriq, u64 ns) > ptp_qoriq->write(®s->ctrl_regs->tmr_cnt_h, hi); } > > +static void tmr_offset_write(struct ptp_qoriq *ptp_qoriq, u64 > +delta_ns) { > + struct ptp_qoriq_registers *regs = &ptp_qoriq->regs; > + u32 hi = delta_ns >> 32; > + u32 lo = delta_ns & 0xffffffff; > + > + ptp_qoriq->write(®s->ctrl_regs->tmroff_l, lo); > + ptp_qoriq->write(®s->ctrl_regs->tmroff_h, hi); } > + > +static u64 tmr_offset_read(struct ptp_qoriq *ptp_qoriq) { > + struct ptp_qoriq_registers *regs = &ptp_qoriq->regs; > + u64 ns; > + u32 lo, hi;
Order variable lines longest to shortest
> + lo = ptp_qoriq->read(®s->ctrl_regs->tmroff_l); > + hi = ptp_qoriq->read(®s->ctrl_regs->tmroff_h); > + ns = ((u64) hi) << 32; > + ns |= lo; > + return ns; > +} > + > /* Caller must hold ptp_qoriq->lock. */ static void set_alarm(struct > ptp_qoriq *ptp_qoriq) { @@ -55,7 +78,9 @@ static void > set_alarm(struct ptp_qoriq *ptp_qoriq) > u64 ns; > u32 lo, hi; > > - ns = tmr_cnt_read(ptp_qoriq) + 1500000000ULL; > + ns = tmr_cnt_read(ptp_qoriq) + tmr_offset_read(ptp_qoriq) > + + 1500000000ULL; > + > ns = div_u64(ns, 1000000000UL) * 1000000000ULL; > ns -= ptp_qoriq->tclk_period; > hi = ns >> 32; > @@ -207,15 +232,12 @@ EXPORT_SYMBOL_GPL(ptp_qoriq_adjfine); > > int ptp_qoriq_adjtime(struct ptp_clock_info *ptp, s64 delta) { > - s64 now; > unsigned long flags; > struct ptp_qoriq *ptp_qoriq = container_of(ptp, struct > ptp_qoriq, caps); > > spin_lock_irqsave(&ptp_qoriq->lock, flags); > > - now = tmr_cnt_read(ptp_qoriq); > - now += delta; > - tmr_cnt_write(ptp_qoriq, now); > + tmr_offset_write(ptp_qoriq, delta);
Writes to the offset register result in an add operation? Or it's a pure write? What will the offset be after a sequence of following adjtime() calls: adjtime(+100); adjtime(+100); adjtime(+100); ? [Nikhil] : It's a pure write operation, I will be sending the updated version of the patch. Wherein retaining the earlier offset value and adding to the new.
> set_fipers(ptp_qoriq); > > spin_unlock_irqrestore(&ptp_qoriq->lock, flags); @@ -232,7 > +254,7 @@ int ptp_qoriq_gettime(struct ptp_clock_info *ptp, struct > timespec64 *ts) > > spin_lock_irqsave(&ptp_qoriq->lock, flags); > > - ns = tmr_cnt_read(ptp_qoriq); > + ns = tmr_cnt_read(ptp_qoriq) + tmr_offset_read(ptp_qoriq); > > spin_unlock_irqrestore(&ptp_qoriq->lock, flags); > > @@ -251,6 +273,8 @@ int ptp_qoriq_settime(struct ptp_clock_info *ptp, > > ns = timespec64_to_ns(ts); > > + tmr_offset_write(ptp_qoriq, 0);
Shouldn't this be under the lock? [Nikhil] : will update this. > spin_lock_irqsave(&ptp_qoriq->lock, flags); > > tmr_cnt_write(ptp_qoriq, ns);
| |