| Message ID | 20200508131338.32956-1-bernard@vivo.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | memory/samsung: reduce unnecessary mutex lock area | expand |
On Fri, May 08, 2020 at 06:13:38AM -0700, Bernard Zhao wrote: > Maybe dmc->df->lock is unnecessary to protect function > exynos5_dmc_perf_events_check(dmc). If we have to protect, > dmc->lock is more better and more effective. > Also, it seems not needed to protect "if (ret) & dev_warn" > branch. > > Signed-off-by: Bernard Zhao <bernard@vivo.com> > --- > drivers/memory/samsung/exynos5422-dmc.c | 6 ++---- > 1 file changed, 2 insertions(+), 4 deletions(-) I checked the concurrent accesses and it looks correct. Lukasz, any review from your side? Best regards, Krzysztof
Hi Krzysztof, I am sorry, I was a bit busy recently. On 5/12/20 7:50 AM, Krzysztof Kozlowski wrote: > On Fri, May 08, 2020 at 06:13:38AM -0700, Bernard Zhao wrote: >> Maybe dmc->df->lock is unnecessary to protect function >> exynos5_dmc_perf_events_check(dmc). If we have to protect, >> dmc->lock is more better and more effective. >> Also, it seems not needed to protect "if (ret) & dev_warn" >> branch. >> >> Signed-off-by: Bernard Zhao <bernard@vivo.com> >> --- >> drivers/memory/samsung/exynos5422-dmc.c | 6 ++---- >> 1 file changed, 2 insertions(+), 4 deletions(-) > > I checked the concurrent accesses and it looks correct. > > Lukasz, any review from your side? The lock from devfreq lock protects from a scenario when concurrent access from devfreq framework uses internal dmc fields 'load' and 'total' (which are set to 'busy_time', 'total_time'). The .get_dev_status can be called at any time (even due to thermal devfreq cooling action) and reads above fields. That's why the calculation of the new values inside dmc is protected. This patch should not be taken IMO. Maybe we can release lock before the if statement, just to speed-up. Regards, Lukasz > > Best regards, > Krzysztof >
On Tue, 12 May 2020 at 10:47, Lukasz Luba <lukasz.luba@arm.com> wrote: > > Hi Krzysztof, > > I am sorry, I was a bit busy recently. > > On 5/12/20 7:50 AM, Krzysztof Kozlowski wrote: > > On Fri, May 08, 2020 at 06:13:38AM -0700, Bernard Zhao wrote: > >> Maybe dmc->df->lock is unnecessary to protect function > >> exynos5_dmc_perf_events_check(dmc). If we have to protect, > >> dmc->lock is more better and more effective. > >> Also, it seems not needed to protect "if (ret) & dev_warn" > >> branch. > >> > >> Signed-off-by: Bernard Zhao <bernard@vivo.com> > >> --- > >> drivers/memory/samsung/exynos5422-dmc.c | 6 ++---- > >> 1 file changed, 2 insertions(+), 4 deletions(-) > > > > I checked the concurrent accesses and it looks correct. > > > > Lukasz, any review from your side? > > The lock from devfreq lock protects from a scenario when > concurrent access from devfreq framework uses internal dmc fields 'load' > and 'total' (which are set to 'busy_time', 'total_time'). > The .get_dev_status can be called at any time (even due to thermal > devfreq cooling action) and reads above fields. > That's why the calculation of the new values inside dmc is protected. I looked at this path (get_dev_status) and currently in devfreq it will be only called from update_devfreq() -> get_target_freq()... at least when looking at devfreq core and governors. On the other hand you are right that this is public function and this call scenario might change. It could be called directly from other paths sooner or later. > This patch should not be taken IMO. Maybe we can release lock before the > if statement, just to speed-up. Yep. Bernard, you can send just this part of the patch. Best regards, Krzysztof
From: Krzysztof Kozlowski <krzk@kernel.org> Date: 2020-05-12 17:05:28 To: Lukasz Luba <lukasz.luba@arm.com> Cc: Bernard Zhao <bernard@vivo.com>,Kukjin Kim <kgene@kernel.org>,linux-pm@vger.kernel.org,"linux-samsung-soc@vger.kernel.org" <linux-samsung-soc@vger.kernel.org>,linux-arm-kernel@lists.infradead.org,"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,opensource.kernel@vivo.com Subject: Re: [PATCH] memory/samsung: reduce unnecessary mutex lock area>On Tue, 12 May 2020 at 10:47, Lukasz Luba <lukasz.luba@arm.com> wrote: >> >> Hi Krzysztof, >> >> I am sorry, I was a bit busy recently. >> >> On 5/12/20 7:50 AM, Krzysztof Kozlowski wrote: >> > On Fri, May 08, 2020 at 06:13:38AM -0700, Bernard Zhao wrote: >> >> Maybe dmc->df->lock is unnecessary to protect function >> >> exynos5_dmc_perf_events_check(dmc). If we have to protect, >> >> dmc->lock is more better and more effective. >> >> Also, it seems not needed to protect "if (ret) & dev_warn" >> >> branch. >> >> >> >> Signed-off-by: Bernard Zhao <bernard@vivo.com> >> >> --- >> >> drivers/memory/samsung/exynos5422-dmc.c | 6 ++---- >> >> 1 file changed, 2 insertions(+), 4 deletions(-) >> > >> > I checked the concurrent accesses and it looks correct. >> > >> > Lukasz, any review from your side? >> >> The lock from devfreq lock protects from a scenario when >> concurrent access from devfreq framework uses internal dmc fields 'load' >> and 'total' (which are set to 'busy_time', 'total_time'). >> The .get_dev_status can be called at any time (even due to thermal >> devfreq cooling action) and reads above fields. >> That's why the calculation of the new values inside dmc is protected. > >I looked at this path (get_dev_status) and currently in devfreq it >will be only called from update_devfreq() -> get_target_freq()... at >least when looking at devfreq core and governors. On the other hand >you are right that this is public function and this call scenario >might change. It could be called directly from other paths sooner or >later. > >> This patch should not be taken IMO. Maybe we can release lock before the >> if statement, just to speed-up. > >Yep. > >Bernard, you can send just this part of the patch. > Sure, I will resubmit a patch in v2. Best regards, Bernard >Best regards, >Krzysztof
On 5/12/20 10:05 AM, Krzysztof Kozlowski wrote: > On Tue, 12 May 2020 at 10:47, Lukasz Luba <lukasz.luba@arm.com> wrote: >> >> Hi Krzysztof, >> >> I am sorry, I was a bit busy recently. >> >> On 5/12/20 7:50 AM, Krzysztof Kozlowski wrote: >>> On Fri, May 08, 2020 at 06:13:38AM -0700, Bernard Zhao wrote: >>>> Maybe dmc->df->lock is unnecessary to protect function >>>> exynos5_dmc_perf_events_check(dmc). If we have to protect, >>>> dmc->lock is more better and more effective. >>>> Also, it seems not needed to protect "if (ret) & dev_warn" >>>> branch. >>>> >>>> Signed-off-by: Bernard Zhao <bernard@vivo.com> >>>> --- >>>> drivers/memory/samsung/exynos5422-dmc.c | 6 ++---- >>>> 1 file changed, 2 insertions(+), 4 deletions(-) >>> >>> I checked the concurrent accesses and it looks correct. >>> >>> Lukasz, any review from your side? >> >> The lock from devfreq lock protects from a scenario when >> concurrent access from devfreq framework uses internal dmc fields 'load' >> and 'total' (which are set to 'busy_time', 'total_time'). >> The .get_dev_status can be called at any time (even due to thermal >> devfreq cooling action) and reads above fields. >> That's why the calculation of the new values inside dmc is protected. > > I looked at this path (get_dev_status) and currently in devfreq it > will be only called from update_devfreq() -> get_target_freq()... at > least when looking at devfreq core and governors. On the other hand > you are right that this is public function and this call scenario > might change. It could be called directly from other paths sooner or > later. Indeed, I am currently changing this while I am adding devfreq devices to the Energy Model. > >> This patch should not be taken IMO. Maybe we can release lock before the >> if statement, just to speed-up. > > Yep. > > Bernard, you can send just this part of the patch. Thank you Bernard and please submit the patch v2. > > Best regards, > Krzysztof > Thank you Krzysztof for your time spent on this. Regards, Lukasz
diff --git a/drivers/memory/samsung/exynos5422-dmc.c b/drivers/memory/samsung/exynos5422-dmc.c index 22a43d662833..88e8ac8b5327 100644 --- a/drivers/memory/samsung/exynos5422-dmc.c +++ b/drivers/memory/samsung/exynos5422-dmc.c @@ -1345,16 +1345,14 @@ static irqreturn_t dmc_irq_thread(int irq, void *priv) int res; struct exynos5_dmc *dmc = priv; - mutex_lock(&dmc->df->lock); - exynos5_dmc_perf_events_check(dmc); + mutex_lock(&dmc->df->lock); res = update_devfreq(dmc->df); + mutex_unlock(&dmc->df->lock); if (res) dev_warn(dmc->dev, "devfreq failed with %d\n", res); - mutex_unlock(&dmc->df->lock); - return IRQ_HANDLED; }
Maybe dmc->df->lock is unnecessary to protect function exynos5_dmc_perf_events_check(dmc). If we have to protect, dmc->lock is more better and more effective. Also, it seems not needed to protect "if (ret) & dev_warn" branch. Signed-off-by: Bernard Zhao <bernard@vivo.com> --- drivers/memory/samsung/exynos5422-dmc.c | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-)