Hi Kevin,
On Fri, 8 Jan 2010, Kevin Hilman wrote:
> First, this patch adds new worst-case latency values to the
> omap_device_pm_latency struct. Here the worst-case measured latencies
> for the activate and deactivate hooks are stored.
>
> In addition, add an option to auto-adjust the latency values used for
> device activate/deactivate.
>
> By setting a new 'OMAP_DEVICE_LATENCY_AUTO_ADJUST' flag in the
> omap_device_pm_latency struct, the omap_device layer automatically
> adjusts the activate/deactivate latencies to the worst-case measured
> values.
>
> Anytime a new worst-case value is found, it is printed to the console.
> Here is an example log during boot using UART2 s an example. After
> boot, the OPP is manually changed to the 125MHz OPP:
>
> [...]
> Freeing init memory: 128K
> omap_device: serial8250.2: new worst case deactivate latency 0: 30517
> omap_device: serial8250.2: new worst case activate latency 0: 30517
> omap_device: serial8250.2: new worst case activate latency 0: 218139648
> omap_device: serial8250.2: new worst case deactivate latency 0: 61035
> omap_device: serial8250.2: new worst case activate latency 0: 278076171
> omap_device: serial8250.2: new worst case activate latency 0: 298614501
> omap_device: serial8250.2: new worst case activate latency 0: 327331542
>
> / # echo 125000 > /sys/devices/system/cpu/cpu0/cpufreq/scaling_setspeed
>
> omap_device: serial8250.2: new worst case deactivate latency 0: 91552
>
> Motivation: this can be used as a technique to automatically determine
> the worst case latency values. The current method of printing a
> warning on every violation is too noisy to actually interact the
> console in order to set low OPP to discover latencies.
>
> Another motivation for this patch is that the activate/deactivate
> latenices can vary depending on the idlemode of the device. While
> working on the UARTs, I noticed that when using no-idle, the activate
> latencies were as high as several hundred msecs as shown above. When
> the UARTs are in smart-idle, the max latency is well under 100 usecs.
>
> Signed-off-by: Kevin Hilman <khilman@xxxxxxxxxxxxxxxxxxx>
> ---
> arch/arm/plat-omap/include/plat/omap_device.h | 4 ++
> arch/arm/plat-omap/omap_device.c | 41 ++++++++++++++++++++-----
> 2 files changed, 37 insertions(+), 8 deletions(-)
>
> diff --git a/arch/arm/plat-omap/include/plat/omap_device.h b/arch/arm/plat-omap/include/plat/omap_device.h
> index dc1fac1..76d4917 100644
> --- a/arch/arm/plat-omap/include/plat/omap_device.h
> +++ b/arch/arm/plat-omap/include/plat/omap_device.h
> @@ -131,11 +131,15 @@ int omap_device_enable_clocks(struct omap_device *od);
> */
> struct omap_device_pm_latency {
> u32 deactivate_lat;
> + u32 deactivate_lat_worst;
> int (*deactivate_func)(struct omap_device *od);
> u32 activate_lat;
> + u32 activate_lat_worst;
> int (*activate_func)(struct omap_device *od);
> + u32 flags;
> };
>
> +#define OMAP_DEVICE_LATENCY_AUTO_ADJUST BIT(1)
>
> /* Get omap_device pointer from platform_device pointer */
> #define to_omap_device(x) container_of((x), struct omap_device, pdev)
> diff --git a/arch/arm/plat-omap/omap_device.c b/arch/arm/plat-omap/omap_device.c
> index 1e5648d..d8c75c8 100644
> --- a/arch/arm/plat-omap/omap_device.c
> +++ b/arch/arm/plat-omap/omap_device.c
> @@ -148,10 +148,22 @@ static int _omap_device_activate(struct omap_device *od, u8 ignore_lat)
> "%llu nsec\n", od->pdev.name, od->pm_lat_level,
> act_lat);
>
> - WARN(act_lat > odpl->activate_lat, "omap_device: %s.%d: "
> - "activate step %d took longer than expected (%llu > %d)\n",
> - od->pdev.name, od->pdev.id, od->pm_lat_level,
> - act_lat, odpl->activate_lat);
> + if (act_lat > odpl->activate_lat) {
> + odpl->activate_lat_worst = act_lat;
> + if (odpl->flags & OMAP_DEVICE_LATENCY_AUTO_ADJUST) {
> + odpl->activate_lat = act_lat;
> + pr_warning("omap_device: %s.%d: new worst case "
> + "activate latency %d: %llu\n",
> + od->pdev.name, od->pdev.id,
> + od->pm_lat_level, act_lat);
> + } else
> + pr_warning("omap_device: %s.%d: activate "
> + "latency %d higher than exptected. "
> + "(%llu > %d)\n",
> + od->pdev.name, od->pdev.id,
> + od->pm_lat_level, act_lat,
> + odpl->activate_lat);
> + }
>
> od->dev_wakeup_lat -= odpl->activate_lat;
> }
> @@ -204,10 +216,23 @@ static int _omap_device_deactivate(struct omap_device *od, u8 ignore_lat)
> "%llu nsec\n", od->pdev.name, od->pm_lat_level,
> deact_lat);
>
> - WARN(deact_lat > odpl->deactivate_lat, "omap_device: %s.%d: "
> - "deactivate step %d took longer than expected "
> - "(%llu > %d)\n", od->pdev.name, od->pdev.id,
> - od->pm_lat_level, deact_lat, odpl->deactivate_lat);
> + if (deact_lat > odpl->deactivate_lat) {
> + odpl->deactivate_lat_worst = deact_lat;
> + if (odpl->flags & OMAP_DEVICE_LATENCY_AUTO_ADJUST) {
> + odpl->deactivate_lat = deact_lat;
> + pr_warning("omap_device: %s.%d: new worst case "
> + "deactivate latency %d: %llu\n",
> + od->pdev.name, od->pdev.id,
> + od->pm_lat_level, deact_lat);
> + } else
> + pr_warning("omap_device: %s.%d: deactivate "
> + "latency %d higher than exptected. "
> + "(%llu > %d)\n",
> + od->pdev.name, od->pdev.id,
> + od->pm_lat_level, deact_lat,
> + odpl->deactivate_lat);
> + }
> +
>
> od->dev_wakeup_lat += odpl->activate_lat;
>
Thanks, looks very useful. The no-idle thing is extremely strange and
useful to know.
Queued for 2.6.34.
- Paul
--
To unsubscribe from this list: send the line "unsubscribe linux-omap" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
