Message ID | 20180516034954.56475-1-code@mmayer.net (mailing list archive) |
---|---|
State | Rejected, archived |
Headers | show |
On 15-05-18, 20:49, Markus Mayer wrote: > From: Markus Mayer <mmayer@broadcom.com> > > Most CPUfreq drivers (at least on ARM) seem to be sorting the available > frequencies from lowest to highest. To match this behaviour, we reverse > the sorting order in brcmstb-avs-cpufreq, so it is now also lowest to > highest. The reasoning isn't correct. Just because everyone else is doing it doesn't make it right and so you shouldn't change just because of that. What you must written instead in the commit log is that the cpufreq core performs better if the table is sorted (in any order), and so we must sort it as well. But I feel the table is already sorted for your platform, isn't it? And I don't see a clear advantage of merging this patch. > Signed-off-by: Markus Mayer <mmayer@broadcom.com> > --- > drivers/cpufreq/brcmstb-avs-cpufreq.c | 9 +++++---- > 1 file changed, 5 insertions(+), 4 deletions(-) > > diff --git a/drivers/cpufreq/brcmstb-avs-cpufreq.c b/drivers/cpufreq/brcmstb-avs-cpufreq.c > index b07559b9ed99..7dac3205d3eb 100644 > --- a/drivers/cpufreq/brcmstb-avs-cpufreq.c > +++ b/drivers/cpufreq/brcmstb-avs-cpufreq.c > @@ -403,7 +403,7 @@ brcm_avs_get_freq_table(struct device *dev, struct private_data *priv) > { > struct cpufreq_frequency_table *table; > unsigned int pstate; > - int i, ret; > + int p, i, ret; > > /* Remember P-state for later */ > ret = brcm_avs_get_pstate(priv, &pstate); > @@ -415,12 +415,13 @@ brcm_avs_get_freq_table(struct device *dev, struct private_data *priv) > if (!table) > return ERR_PTR(-ENOMEM); > > - for (i = AVS_PSTATE_P0; i <= AVS_PSTATE_MAX; i++) { > - ret = brcm_avs_set_pstate(priv, i); > + for (p = AVS_PSTATE_MAX, i = 0; p >= 0; p--, i++) { > + ret = brcm_avs_set_pstate(priv, p); > if (ret) > return ERR_PTR(ret); > table[i].frequency = brcm_avs_get_frequency(priv->base); > - table[i].driver_data = i; > + /* Store the corresponding P-state with each frequency */ > + table[i].driver_data = p; > } > table[i].frequency = CPUFREQ_TABLE_END; > > -- > 2.7.4
On 05/15/2018 09:32 PM, Viresh Kumar wrote: > On 15-05-18, 20:49, Markus Mayer wrote: >> From: Markus Mayer <mmayer@broadcom.com> >> >> Most CPUfreq drivers (at least on ARM) seem to be sorting the available >> frequencies from lowest to highest. To match this behaviour, we reverse >> the sorting order in brcmstb-avs-cpufreq, so it is now also lowest to >> highest. > > The reasoning isn't correct. Just because everyone else is doing it > doesn't make it right and so you shouldn't change just because of > that. > > What you must written instead in the commit log is that the cpufreq > core performs better if the table is sorted (in any order), and so we > must sort it as well. Is there a reason why set_freq_table_sorted() tries an ascending or descending sort, but does not enforce one versus another for all drivers? > > But I feel the table is already sorted for your platform, isn't it? > And I don't see a clear advantage of merging this patch. The patch changes the order to have the lowest to highest, whereas the current implementation has them from highest to lowest. From what you are saying, it sounds like this is unnecessary, since the sorting is already making things efficient enough, so this is just a cosmetic thing? > >> Signed-off-by: Markus Mayer <mmayer@broadcom.com> >> --- >> drivers/cpufreq/brcmstb-avs-cpufreq.c | 9 +++++---- >> 1 file changed, 5 insertions(+), 4 deletions(-) >> >> diff --git a/drivers/cpufreq/brcmstb-avs-cpufreq.c b/drivers/cpufreq/brcmstb-avs-cpufreq.c >> index b07559b9ed99..7dac3205d3eb 100644 >> --- a/drivers/cpufreq/brcmstb-avs-cpufreq.c >> +++ b/drivers/cpufreq/brcmstb-avs-cpufreq.c >> @@ -403,7 +403,7 @@ brcm_avs_get_freq_table(struct device *dev, struct private_data *priv) >> { >> struct cpufreq_frequency_table *table; >> unsigned int pstate; >> - int i, ret; >> + int p, i, ret; >> >> /* Remember P-state for later */ >> ret = brcm_avs_get_pstate(priv, &pstate); >> @@ -415,12 +415,13 @@ brcm_avs_get_freq_table(struct device *dev, struct private_data *priv) >> if (!table) >> return ERR_PTR(-ENOMEM); >> >> - for (i = AVS_PSTATE_P0; i <= AVS_PSTATE_MAX; i++) { >> - ret = brcm_avs_set_pstate(priv, i); >> + for (p = AVS_PSTATE_MAX, i = 0; p >= 0; p--, i++) { >> + ret = brcm_avs_set_pstate(priv, p); >> if (ret) >> return ERR_PTR(ret); >> table[i].frequency = brcm_avs_get_frequency(priv->base); >> - table[i].driver_data = i; >> + /* Store the corresponding P-state with each frequency */ >> + table[i].driver_data = p; >> } >> table[i].frequency = CPUFREQ_TABLE_END; >> >> -- >> 2.7.4 >
On 16-05-18, 12:24, Florian Fainelli wrote: > On 05/15/2018 09:32 PM, Viresh Kumar wrote: > > On 15-05-18, 20:49, Markus Mayer wrote: > >> From: Markus Mayer <mmayer@broadcom.com> > >> > >> Most CPUfreq drivers (at least on ARM) seem to be sorting the available > >> frequencies from lowest to highest. To match this behaviour, we reverse > >> the sorting order in brcmstb-avs-cpufreq, so it is now also lowest to > >> highest. > > > > The reasoning isn't correct. Just because everyone else is doing it > > doesn't make it right and so you shouldn't change just because of > > that. > > > > What you must written instead in the commit log is that the cpufreq > > core performs better if the table is sorted (in any order), and so we > > must sort it as well. > > Is there a reason why set_freq_table_sorted() tries an ascending or > descending sort, but does not enforce one versus another for all drivers? set_freq_table_sorted() doesn't sort the frequency table but checks if the table is already sorted in a particular order. And then cpufreq_frequency_table_target() is optimized based on this flag. We don't have to enforce any particular ordering here. > > But I feel the table is already sorted for your platform, isn't it? > > And I don't see a clear advantage of merging this patch. > > The patch changes the order to have the lowest to highest, whereas the > current implementation has them from highest to lowest. From what you > are saying, it sounds like this is unnecessary, since the sorting is > already making things efficient enough, so this is just a cosmetic thing? Right. It shouldn't make any performance improvements.
diff --git a/drivers/cpufreq/brcmstb-avs-cpufreq.c b/drivers/cpufreq/brcmstb-avs-cpufreq.c index b07559b9ed99..7dac3205d3eb 100644 --- a/drivers/cpufreq/brcmstb-avs-cpufreq.c +++ b/drivers/cpufreq/brcmstb-avs-cpufreq.c @@ -403,7 +403,7 @@ brcm_avs_get_freq_table(struct device *dev, struct private_data *priv) { struct cpufreq_frequency_table *table; unsigned int pstate; - int i, ret; + int p, i, ret; /* Remember P-state for later */ ret = brcm_avs_get_pstate(priv, &pstate); @@ -415,12 +415,13 @@ brcm_avs_get_freq_table(struct device *dev, struct private_data *priv) if (!table) return ERR_PTR(-ENOMEM); - for (i = AVS_PSTATE_P0; i <= AVS_PSTATE_MAX; i++) { - ret = brcm_avs_set_pstate(priv, i); + for (p = AVS_PSTATE_MAX, i = 0; p >= 0; p--, i++) { + ret = brcm_avs_set_pstate(priv, p); if (ret) return ERR_PTR(ret); table[i].frequency = brcm_avs_get_frequency(priv->base); - table[i].driver_data = i; + /* Store the corresponding P-state with each frequency */ + table[i].driver_data = p; } table[i].frequency = CPUFREQ_TABLE_END;