[PATCH] hwrng: bcm2835: Fix hwrng throughput regression

[Jason A. Donenfeld]

On Wed, Aug 30, 2023 at 08:05:39PM +0200, Stefan Wahren wrote:
> Hi Jason,
> 
> Am 26.08.23 um 17:48 schrieb Jason A. Donenfeld:
> > On Sat, Aug 26, 2023 at 04:01:58PM +0200, Stefan Wahren wrote:
> >> Hi Jason,
> >>
> >> Am 26.08.23 um 14:34 schrieb Jason A. Donenfeld:
> >>> On Sat, Aug 26, 2023 at 01:28:28PM +0200, Stefan Wahren wrote:
> >>>> The recent RCU stall fix caused a massive throughput regression of the
> >>>> hwrng on Raspberry Pi 0 - 3. So try to restore a similiar throughput
> >>>> as before the RCU stall fix.
> >>>>
> >>>> Some performance measurements on Raspberry Pi 3B+ (arm64/defconfig):
> >>>>
> >>>> sudo dd if=/dev/hwrng of=/dev/null count=1 bs=10000
> >>>>
> >>>> cpu_relax              ~138025 Bytes / sec
> >>>> hwrng_msleep(1000)         ~13 Bytes / sec
> >>>> usleep_range(100,200)   ~92141 Bytes / sec
> >>>>
> >>>> Fixes: 96cb9d055445 ("hwrng: bcm2835 - use hwrng_msleep() instead of cpu_relax()")
> >>>> Link: https://lore.kernel.org/linux-arm-kernel/bc97ece5-44a3-4c4e-77da-2db3eb66b128@gmx.net/
> >>>> Signed-off-by: Stefan Wahren <wahrenst at gmx.net>
> >>>> ---
> >>>>    drivers/char/hw_random/bcm2835-rng.c | 3 ++-
> >>>>    1 file changed, 2 insertions(+), 1 deletion(-)
> >>>>
> >>>> diff --git a/drivers/char/hw_random/bcm2835-rng.c b/drivers/char/hw_random/bcm2835-rng.c
> >>>> index e98fcac578d6..3f1b6aaa98ee 100644
> >>>> --- a/drivers/char/hw_random/bcm2835-rng.c
> >>>> +++ b/drivers/char/hw_random/bcm2835-rng.c
> >>>> @@ -14,6 +14,7 @@
> >>>>    #include <linux/printk.h>
> >>>>    #include <linux/clk.h>
> >>>>    #include <linux/reset.h>
> >>>> +#include <linux/delay.h>
> >>>>
> >>>>    #define RNG_CTRL	0x0
> >>>>    #define RNG_STATUS	0x4
> >>>> @@ -71,7 +72,7 @@ static int bcm2835_rng_read(struct hwrng *rng, void *buf, size_t max,
> >>>>    	while ((rng_readl(priv, RNG_STATUS) >> 24) == 0) {
> >>>>    		if (!wait)
> >>>>    			return 0;
> >>>> -		hwrng_msleep(rng, 1000);
> >>>> +		usleep_range(100, 200);
> >>> I think we still need to use the hwrng_msleep function so that the sleep
> >>> remains cancelable. Maybe just change the 1000 to 100?
> >> i found that other hwrng driver like iproc-rng200 (Raspberry Pi 4) also
> >> use usleep_range().
> >>
> >> Nevertheless here are more numbers:
> >>
> >> usleep_range(200,400) : 47776 bytes / sec
> >> hwrng_msleep(20) : 715 bytes / sec
> >>
> >> Changing to 100 ms won't be a real gain.
> > I'm fine with whatever number you want there. Maybe we need a
> > hwrng_usleep_range() that takes into account rng->dying like
> > hwrng_msleep() does? (And iproc-rng200 should probably use that too?)
> the idea of this patch was to fix the performance regression in upcoming
> mainline and backport the fix to Linux 6.1 LTS. After that i'm fine with
> the introduction of hwrng_usleep_range().

No, you've got it backwards, but maybe it's a bit confusing why that is.

Originally, lots of drivers called variants of the msleep/usleep
functions. But this lead to some subtle bugs when the hwrng subsystem
was changing the active RNG or when the driver was unloading or during
suspend or a couple of other edge cases. In some cases, I believe there
were some interesting deadlocks. The solution to it was to make those
sleep calls cancellable with the &rng->dying completion, so those sleeps
could break, regardless of the time or the state of timekeeping on the
system. The &rng->dying thing was encapsulated in the hwrng_msleep()
function, and now it's a rule that any sleeps in hwrng drivers have got
to go through that.

Now, it sounds like you've come up with a case where hwrng_msleep(1) is
too long of a sleep for your purposes. In that case, indeed a usleep
variant might be called for. But in doing so, don't _reintroduce the
same bug we already fixed_. In other words, don't fix one regression by
adding another. Instead, follow the established pattern, and add a
hwrng_usleep_whatever() that waits on that same &rng->dying.

Jason