2023-08-26 13:59:18

by Liao, Chang

Subject: [PATCH] cpufreq: pcc: Fix the potentinal scheduling delays in target_index()

pcc_cpufreq_target():
cpufreq_freq_transition_begin();
spin_lock(&pcc_lock);
[critical section]
cpufreq_freq_transition_end();
spin_unlock(&pcc_lock);

Above code has a performance issue, consider that Task0 executes
'cpufreq_freq_transition_end()' to wake Task1 and preempted imediatedly
without releasing 'pcc_lock', then Task1 needs to wait for Task0 to
release 'pcc_lock'. In the worst case, this locking order can result in
Task1 wasting two scheduling rounds before it can enter the critical
section.

Signed-off-by: Liao Chang <[email protected]>
---
drivers/cpufreq/pcc-cpufreq.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/cpufreq/pcc-cpufreq.c b/drivers/cpufreq/pcc-cpufreq.c
index 73efbcf5513b..9d732a00e2a5 100644
--- a/drivers/cpufreq/pcc-cpufreq.c
+++ b/drivers/cpufreq/pcc-cpufreq.c
@@ -232,8 +232,8 @@ static int pcc_cpufreq_target(struct cpufreq_policy *policy,
status = ioread16(&pcch_hdr->status);
iowrite16(0, &pcch_hdr->status);

- cpufreq_freq_transition_end(policy, &freqs, status != CMD_COMPLETE);
spin_unlock(&pcc_lock);
+ cpufreq_freq_transition_end(policy, &freqs, status != CMD_COMPLETE);

if (status != CMD_COMPLETE) {
pr_debug("target: FAILED for cpu %d, with status: 0x%x\n",
--
2.34.1

2023-08-29 19:16:40

by Rafael J. Wysocki

Subject: Re: [PATCH] cpufreq: pcc: Fix the potentinal scheduling delays in target_index()

On Mon, Aug 28, 2023 at 9:02 AM Viresh Kumar <[email protected]> wrote:
>
> On 26-08-23, 09:57, Liao Chang wrote:
> > pcc_cpufreq_target():
> > cpufreq_freq_transition_begin();
> > spin_lock(&pcc_lock);
> > [critical section]
> > cpufreq_freq_transition_end();
> > spin_unlock(&pcc_lock);
> >
> > Above code has a performance issue, consider that Task0 executes
> > 'cpufreq_freq_transition_end()' to wake Task1 and preempted imediatedly
> > without releasing 'pcc_lock', then Task1 needs to wait for Task0 to
> > release 'pcc_lock'. In the worst case, this locking order can result in
> > Task1 wasting two scheduling rounds before it can enter the critical
> > section.
> >
> > Signed-off-by: Liao Chang <[email protected]>
> > ---
> > drivers/cpufreq/pcc-cpufreq.c | 2 +-
> > 1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/drivers/cpufreq/pcc-cpufreq.c b/drivers/cpufreq/pcc-cpufreq.c
> > index 73efbcf5513b..9d732a00e2a5 100644
> > --- a/drivers/cpufreq/pcc-cpufreq.c
> > +++ b/drivers/cpufreq/pcc-cpufreq.c
> > @@ -232,8 +232,8 @@ static int pcc_cpufreq_target(struct cpufreq_policy *policy,
> > status = ioread16(&pcch_hdr->status);
> > iowrite16(0, &pcch_hdr->status);
> >
> > - cpufreq_freq_transition_end(policy, &freqs, status != CMD_COMPLETE);
> > spin_unlock(&pcc_lock);
> > + cpufreq_freq_transition_end(policy, &freqs, status != CMD_COMPLETE);
> >
> > if (status != CMD_COMPLETE) {
> > pr_debug("target: FAILED for cpu %d, with status: 0x%x\n",
>
> Acked-by: Viresh Kumar <[email protected]>

Applied as 6.6-rc material, thanks!