Re: [PATCH] PCI: aardvark: Don't rely on jiffies while holding spinlock

From: Remi Pommarel
Date: Fri Sep 27 2019 - 04:16:50 EST


Hi Thomas,

Thanks for the review.

On Wed, Sep 25, 2019 at 11:33:51AM +0200, Thomas Petazzoni wrote:
> Hello Remi,
>
> Thanks for the patch, I have a few comments/questions below.
>
> On Sun, 1 Sep 2019 16:23:03 +0200
> Remi Pommarel <repk@xxxxxxxxxxxx> wrote:
>
> > diff --git a/drivers/pci/controller/pci-aardvark.c b/drivers/pci/controller/pci-aardvark.c
> > index fc0fe4d4de49..1fa6d04ad7aa 100644
> > --- a/drivers/pci/controller/pci-aardvark.c
> > +++ b/drivers/pci/controller/pci-aardvark.c
> > @@ -175,7 +175,8 @@
> > (PCIE_CONF_BUS(bus) | PCIE_CONF_DEV(PCI_SLOT(devfn)) | \
> > PCIE_CONF_FUNC(PCI_FUNC(devfn)) | PCIE_CONF_REG(where))
> >
> > -#define PIO_TIMEOUT_MS 1
> > +#define PIO_RETRY_CNT 10
> > +#define PIO_RETRY_DELAY 100 /* 100 us*/
> >
> > #define LINK_WAIT_MAX_RETRIES 10
> > #define LINK_WAIT_USLEEP_MIN 90000
> > @@ -383,17 +384,16 @@ static void advk_pcie_check_pio_status(struct advk_pcie *pcie)
> > static int advk_pcie_wait_pio(struct advk_pcie *pcie)
> > {
> > struct device *dev = &pcie->pdev->dev;
> > - unsigned long timeout;
> > + size_t i;
>
> Is it common to use a size_t for a loop counter ?

It was for me but seem not to be used that much. I can change that to an
int.

> >
> > - timeout = jiffies + msecs_to_jiffies(PIO_TIMEOUT_MS);
> > -
> > - while (time_before(jiffies, timeout)) {
> > + for (i = 0; i < PIO_RETRY_CNT; ++i) {
>
> I find it more common to use post-increment for loop counters rather
> than pre-increment, but that's a really nitpick and I don't care much.
>

Will change that to post-increment.

> > u32 start, isr;
> >
> > start = advk_readl(pcie, PIO_START);
> > isr = advk_readl(pcie, PIO_ISR);
> > if (!start && isr)
> > return 0;
> > + udelay(PIO_RETRY_DELAY);
>
> But the bigger issue is that this change causes a 100us delay at
> *every* single PIO read or write operation.
>
> Indeed, at the first iteration of the loop, the PIO operation has not
> completed, so you will always hit the udelay(100) a first time, and
> it's only at the second iteration of the loop that the PIO operation
> has completed (for successful PIO operations of course, which don't hit
> the timeout).
>
> I took a measurement around wait_pio() with sched_clock before and
> after the patch. Before the patch, I have measurements like this (in
> nanoseconds):
>
> [ 1.562801] time = 6000
> [ 1.565310] time = 6000
> [ 1.567809] time = 6080
> [ 1.570327] time = 6080
> [ 1.572836] time = 6080
> [ 1.575339] time = 6080
> [ 1.577858] time = 2720
> [ 1.580366] time = 2720
> [ 1.582862] time = 6000
> [ 1.585377] time = 2720
> [ 1.587890] time = 2720
> [ 1.590393] time = 2720
>
> So it takes a few microseconds for each PIO operation.
>
> With your patch applied:
>
> [ 2.267291] time = 101680
> [ 2.270002] time = 100880
> [ 2.272852] time = 100800
> [ 2.275573] time = 100880
> [ 2.278285] time = 100800
> [ 2.281005] time = 100880
> [ 2.283722] time = 100800
> [ 2.286444] time = 100880
> [ 2.289264] time = 100880
> [ 2.291981] time = 100800
> [ 2.294690] time = 100800
> [ 2.297405] time = 100800
>
> We're jumping to 100us for every PIO read/write operation. To be
> honest, I don't know if this is very important, there are not that many
> PIO operations, and they are not used in any performance hot path. But
> I thought it was worth pointing out the additional delay caused by this
> implementation change.

Good catch thanks for the measurements, will move to a 2us delay.

--
Remi