| Message ID | 20240703110741.2668800-1-quic_sibis@quicinc.com |
|---|---|
| State | New |
| Headers | show |
| Series | pmdomain: arm: Fix debugfs node creation failure | expand |
On 7/5/24 18:34, Sudeep Holla wrote: > On Fri, Jul 05, 2024 at 09:16:29AM +0530, Sibi Sankar wrote: >> >> On 7/4/24 16:02, Sudeep Holla wrote: >>> >>> If there are 2 perf domains for a device or group of devices, there must >>> be something unique about each of these domains. Why can't the firmware >>> specify the uniqueness or the difference via the name? >>> >>> The example above seems firmware is being just lazy to update it. Also >>> for the user/developer/debugger, the unique name might be more useful >>> than just this number. >>> >>> So please use the name(we must now have extended name if 16bytes are less) >>> to provide unique names. Please stop working around such silly firmware >>> bugs like this, it just makes using debugfs for anything useful harder. >> >> This is just meant to address firmware that are already out in the wild. >> That being said I don't necessarily agree with the patch either since >> it's penalizing firmware that actually uses a proper name by appending >> something inherently less useful to it. Since, the using of an unique >> domain name isn't required by the spec, the need for it goes under the radar >> for vendors. Mandating it might be the right thing to do since >> the kernel seems inherently expect that. >> > > Well I would love if spec authors can agree and mandate this. But this is > one of those things I can't argue as I don't necessarily agree with the > argument. There are 2 distinct/unique domains but firmware authors ran out > of unique names for them or just can't be bothered to care about it. > > They can't run out of characters as well in above examples, firmware can > add some useless domain ID in the name if they can't be bothered or creative. > > So I must admit I can't be bothered as well with that honestly. Okay, I guess the conclusion is that if the firmware vendors don't care enough to provide unique names, they get to live without those debugfs nodes. Do we really want to register/expose scmi perf power-domains used by the CPU nodes? Given that scmi-cpufreq doesn't consume these power domains and can be voted upon by another consumer, wouldn't this cause a disconnect? -Sibi > -- > Regards, > Sudeep
> Subject: Re: [PATCH] pmdomain: arm: Fix debugfs node creation failure > > > > On 7/5/24 18:34, Sudeep Holla wrote: > > On Fri, Jul 05, 2024 at 09:16:29AM +0530, Sibi Sankar wrote: > >> > >> On 7/4/24 16:02, Sudeep Holla wrote: > >>> > >>> If there are 2 perf domains for a device or group of devices, there > >>> must be something unique about each of these domains. Why > can't the > >>> firmware specify the uniqueness or the difference via the name? > >>> > >>> The example above seems firmware is being just lazy to update it. > >>> Also for the user/developer/debugger, the unique name might be > more > >>> useful than just this number. > >>> > >>> So please use the name(we must now have extended name if > 16bytes are > >>> less) to provide unique names. Please stop working around such > silly > >>> firmware bugs like this, it just makes using debugfs for anything > useful harder. > >> > >> This is just meant to address firmware that are already out in the > wild. > >> That being said I don't necessarily agree with the patch either since > >> it's penalizing firmware that actually uses a proper name by > >> appending something inherently less useful to it. Since, the using of > >> an unique domain name isn't required by the spec, the need for it > >> goes under the radar for vendors. Mandating it might be the right > >> thing to do since the kernel seems inherently expect that. > >> > > > > Well I would love if spec authors can agree and mandate this. But this > > is one of those things I can't argue as I don't necessarily agree with > > the argument. There are 2 distinct/unique domains but firmware > authors > > ran out of unique names for them or just can't be bothered to care > about it. > > > > They can't run out of characters as well in above examples, firmware > > can add some useless domain ID in the name if they can't be > bothered or creative. As Sibi raised, Spec does not has restriction on name. Linux chose to use genpd to support perf domain, but now it turns out that Linux is forcing firmware to use different name for power/perf domain. This will not convince firmware developers. For example, firmware might be as below: struct pd_perf_domain { char *name; (int *)power_hook(int id); (int *)perf_hook(int level); }; From firmware developer's view, name is shared for pd and perf. The fix should be in linux side. > > > > So I must admit I can't be bothered as well with that honestly. > > Okay, I guess the conclusion is that if the firmware vendors don't care > enough to provide unique names, they get to live without those > debugfs nodes. > > Do we really want to register/expose scmi perf power-domains used by > the CPU nodes? How about not register debugfs for perf? Given that scmi-cpufreq doesn't consume these power > domains and can be voted upon by another consumer, wouldn't this > cause a disconnect? You might be also interested in [1], which is also scmi cpufreq related. [1]https://lore.kernel.org/all/20240729070325.2065286-1-peng.fan@oss.nxp.com/ Regards, Peng. > > -Sibi > > > -- > > Regards, > > Sudeep
+ Peng On Fri, 5 Jul 2024 at 15:04, Sudeep Holla <sudeep.holla@arm.com> wrote: > > On Fri, Jul 05, 2024 at 09:16:29AM +0530, Sibi Sankar wrote: > > > > On 7/4/24 16:02, Sudeep Holla wrote: > > > > > > If there are 2 perf domains for a device or group of devices, there must > > > be something unique about each of these domains. Why can't the firmware > > > specify the uniqueness or the difference via the name? > > > > > > The example above seems firmware is being just lazy to update it. Also > > > for the user/developer/debugger, the unique name might be more useful > > > than just this number. > > > > > > So please use the name(we must now have extended name if 16bytes are less) > > > to provide unique names. Please stop working around such silly firmware > > > bugs like this, it just makes using debugfs for anything useful harder. > > > > This is just meant to address firmware that are already out in the wild. > > That being said I don't necessarily agree with the patch either since > > it's penalizing firmware that actually uses a proper name by appending > > something inherently less useful to it. Since, the using of an unique > > domain name isn't required by the spec, the need for it goes under the radar > > for vendors. Mandating it might be the right thing to do since > > the kernel seems inherently expect that. > > > > Well I would love if spec authors can agree and mandate this. But this is > one of those things I can't argue as I don't necessarily agree with the > argument. There are 2 distinct/unique domains but firmware authors ran out > of unique names for them or just can't be bothered to care about it. > > They can't run out of characters as well in above examples, firmware can > add some useless domain ID in the name if they can't be bothered or creative. > > So I must admit I can't be bothered as well with that honestly. Sudeep, while I understand your point and I agree with it, it's really a simple fix that $subject patch is proposing. As the unique name isn't mandated by the SCMI spec, it looks to me that we should make a fix for it on the Linux side. I have therefore decided to queue up $subject patch for fixes. Please let me know if you have any other proposals/objections moving forward. Kind regards Uffe
On Wed, Aug 14, 2024 at 02:38:24PM +0200, Ulf Hansson wrote: > > Sudeep, while I understand your point and I agree with it, it's really > a simple fix that $subject patch is proposing. As the unique name > isn't mandated by the SCMI spec, it looks to me that we should make a > fix for it on the Linux side. > Yes, I did come to the conclusion that this is inevitable but hadn't thought much on the exact solution. This email and you merging the original patch made me think a bit quickly now 😉 > I have therefore decided to queue up $subject patch for fixes. Please > let me know if you have any other proposals/objections moving forward. The original patch may not work well with the use case Peng presented. As the name and id may also match in their case, I was wondering if we need to add some prefix like perf- or something to avoid the potential clash across power and perf genpds ? I may be missing something still as it is hard to visualise all possible case that can happen with variety of platform and their firmware. In short, happy to have some fix for the issue in some form whichever works for wider set of platforms. -- Regards, Sudeep
On Wed, 14 Aug 2024 at 15:31, Sudeep Holla <sudeep.holla@arm.com> wrote: > > On Wed, Aug 14, 2024 at 02:38:24PM +0200, Ulf Hansson wrote: > > > > Sudeep, while I understand your point and I agree with it, it's really > > a simple fix that $subject patch is proposing. As the unique name > > isn't mandated by the SCMI spec, it looks to me that we should make a > > fix for it on the Linux side. > > > > Yes, I did come to the conclusion that this is inevitable but hadn't > thought much on the exact solution. This email and you merging the original > patch made me think a bit quickly now 😉 Alright, great! > > > I have therefore decided to queue up $subject patch for fixes. Please > > let me know if you have any other proposals/objections moving forward. > > The original patch may not work well with the use case Peng presented. > As the name and id may also match in their case, I was wondering if we > need to add some prefix like perf- or something to avoid the potential > clash across power and perf genpds ? I may be missing something still as > it is hard to visualise all possible case that can happen with variety > of platform and their firmware. > > In short, happy to have some fix for the issue in some form whichever > works for wider set of platforms. Okay, so I have dropped the $subject patch from my fixes branch for now, to allow us and Sibi to come up with an improved approach. That said, it looks to me that the proper fix needs to involve pm_genpd_init() in some way, as this problem with unique device naming isn't really limited to SCMI. Normally we use an "ida" to get a unique index that we tag on to the device's name, but maybe there is a better strategy here!? Kind regards Uffe
On Thu, Aug 15, 2024 at 12:46:15PM +0200, Ulf Hansson wrote: > On Wed, 14 Aug 2024 at 15:31, Sudeep Holla <sudeep.holla@arm.com> wrote: > > > > On Wed, Aug 14, 2024 at 02:38:24PM +0200, Ulf Hansson wrote: > > > > > > Sudeep, while I understand your point and I agree with it, it's really > > > a simple fix that $subject patch is proposing. As the unique name > > > isn't mandated by the SCMI spec, it looks to me that we should make a > > > fix for it on the Linux side. > > > > > > > Yes, I did come to the conclusion that this is inevitable but hadn't > > thought much on the exact solution. This email and you merging the original > > patch made me think a bit quickly now 😉 > > Alright, great! > > > > > > I have therefore decided to queue up $subject patch for fixes. Please > > > let me know if you have any other proposals/objections moving forward. > > > > The original patch may not work well with the use case Peng presented. > > As the name and id may also match in their case, I was wondering if we > > need to add some prefix like perf- or something to avoid the potential > > clash across power and perf genpds ? I may be missing something still as > > it is hard to visualise all possible case that can happen with variety > > of platform and their firmware. > > > > In short, happy to have some fix for the issue in some form whichever > > works for wider set of platforms. > > Okay, so I have dropped the $subject patch from my fixes branch for > now, to allow us and Sibi to come up with an improved approach. > > That said, it looks to me that the proper fix needs to involve > pm_genpd_init() in some way, as this problem with unique device naming > isn't really limited to SCMI. Normally we use an "ida" to get a unique > index that we tag on to the device's name, but maybe there is a better > strategy here!? Yes using "ida" for unique index might work here as well AFAIU. It can be one of the possible solution for sure.
On 8/15/24 19:16, Sudeep Holla wrote: > On Thu, Aug 15, 2024 at 12:46:15PM +0200, Ulf Hansson wrote: >> On Wed, 14 Aug 2024 at 15:31, Sudeep Holla <sudeep.holla@arm.com> wrote: >>> >>> On Wed, Aug 14, 2024 at 02:38:24PM +0200, Ulf Hansson wrote: >>>> >>>> Sudeep, while I understand your point and I agree with it, it's really >>>> a simple fix that $subject patch is proposing. As the unique name >>>> isn't mandated by the SCMI spec, it looks to me that we should make a >>>> fix for it on the Linux side. >>>> >>> >>> Yes, I did come to the conclusion that this is inevitable but hadn't >>> thought much on the exact solution. This email and you merging the original >>> patch made me think a bit quickly now 😉 >> >> Alright, great! >> >>> >>>> I have therefore decided to queue up $subject patch for fixes. Please >>>> let me know if you have any other proposals/objections moving forward. >>> >>> The original patch may not work well with the use case Peng presented. >>> As the name and id may also match in their case, I was wondering if we >>> need to add some prefix like perf- or something to avoid the potential >>> clash across power and perf genpds ? I may be missing something still as >>> it is hard to visualise all possible case that can happen with variety >>> of platform and their firmware. >>> >>> In short, happy to have some fix for the issue in some form whichever >>> works for wider set of platforms. >> >> Okay, so I have dropped the $subject patch from my fixes branch for >> now, to allow us and Sibi to come up with an improved approach. >> >> That said, it looks to me that the proper fix needs to involve >> pm_genpd_init() in some way, as this problem with unique device naming >> isn't really limited to SCMI. Normally we use an "ida" to get a unique >> index that we tag on to the device's name, but maybe there is a better >> strategy here!? > > Yes using "ida" for unique index might work here as well AFAIU. It can be > one of the possible solution for sure. Just re-spun it with ida, I've also shared how the output looks with those additional device ids added to the device name. Have a look at it when you get time. -Sibi >
diff --git a/drivers/pmdomain/arm/scmi_perf_domain.c b/drivers/pmdomain/arm/scmi_perf_domain.c index d7ef46ccd9b8..0af5dc941349 100644 --- a/drivers/pmdomain/arm/scmi_perf_domain.c +++ b/drivers/pmdomain/arm/scmi_perf_domain.c @@ -18,6 +18,7 @@ struct scmi_perf_domain { const struct scmi_perf_proto_ops *perf_ops; const struct scmi_protocol_handle *ph; const struct scmi_perf_domain_info *info; + char domain_name[SCMI_MAX_STR_SIZE]; u32 domain_id; }; @@ -123,7 +124,12 @@ static int scmi_perf_domain_probe(struct scmi_device *sdev) scmi_pd->domain_id = i; scmi_pd->perf_ops = perf_ops; scmi_pd->ph = ph; - scmi_pd->genpd.name = scmi_pd->info->name; + + /* Domain attributes can report identical names across domains */ + snprintf(scmi_pd->domain_name, sizeof(scmi_pd->domain_name), "%s-%d", + scmi_pd->info->name, scmi_pd->domain_id); + + scmi_pd->genpd.name = scmi_pd->domain_name; scmi_pd->genpd.flags = GENPD_FLAG_ALWAYS_ON | GENPD_FLAG_OPP_TABLE_FW; scmi_pd->genpd.set_performance_state = scmi_pd_set_perf_state;
The domain attributes returned by the perf protocol can end up reporting identical names across domains, resulting in debugfs node creation failure. Fix this duplication by appending the domain-id to the domain name. Logs: debugfs: Directory 'NCC' with parent 'pm_genpd' already present! debugfs: Directory 'NCC' with parent 'pm_genpd' already present! Fixes: 2af23ceb8624 ("pmdomain: arm: Add the SCMI performance domain") Signed-off-by: Sibi Sankar <quic_sibis@quicinc.com> --- drivers/pmdomain/arm/scmi_perf_domain.c | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-)