| Message ID | 20241109044151.29804-12-mario.limonciello@amd.com |
|---|---|
| State | Superseded |
| Headers | show |
| Series | Add support for binding ACPI platform profile to multiple drivers | expand |
Am 09.11.24 um 05:41 schrieb Mario Limonciello: > The name attribute shows the name of the associated platform profile > handler. > > Tested-by: Mark Pearson <mpearson-lenovo@squebb.ca> > Signed-off-by: Mario Limonciello <mario.limonciello@amd.com> > --- > drivers/acpi/platform_profile.c | 27 +++++++++++++++++++++++++++ > 1 file changed, 27 insertions(+) > > diff --git a/drivers/acpi/platform_profile.c b/drivers/acpi/platform_profile.c > index ef6af2c655524..4e2eda18f7f5f 100644 > --- a/drivers/acpi/platform_profile.c > +++ b/drivers/acpi/platform_profile.c > @@ -25,8 +25,35 @@ static_assert(ARRAY_SIZE(profile_names) == PLATFORM_PROFILE_LAST); > > static DEFINE_IDA(platform_profile_ida); > > +/** > + * name_show - Show the name of the profile handler > + * @dev: The device > + * @attr: The attribute > + * @buf: The buffer to write to > + * Return: The number of bytes written > + */ > +static ssize_t name_show(struct device *dev, > + struct device_attribute *attr, > + char *buf) > +{ > + struct platform_profile_handler *handler = dev_get_drvdata(dev); > + > + scoped_cond_guard(mutex_intr, return -ERESTARTSYS, &profile_lock) { > + return sysfs_emit(buf, "%s\n", handler->name); > + } > + return -ERESTARTSYS; I still have a bad feeling about the locking inside the class attributes... Can we assume that no sysfs accesses occur after unregistering the class device? Even if this is not the case then the locking fails to protect the platform_profile_handler here. If the device is unregistered right after dev_get_drvdata() was called, then we would sill operate on possibly stale data once we take the profile_lock. Does someone have any clue how sysfs attributes act during removal? Thanks, Armin Wolf > +} > + > +static DEVICE_ATTR_RO(name); > +static struct attribute *profile_attrs[] = { > + &dev_attr_name.attr, > + NULL > +}; > +ATTRIBUTE_GROUPS(profile); > + > static const struct class platform_profile_class = { > .name = "platform-profile", > + .dev_groups = profile_groups, > }; > > static ssize_t platform_profile_choices_show(struct device *dev,
Am 18.11.24 um 20:43 schrieb Armin Wolf: > Am 09.11.24 um 05:41 schrieb Mario Limonciello: > >> The name attribute shows the name of the associated platform profile >> handler. >> >> Tested-by: Mark Pearson <mpearson-lenovo@squebb.ca> >> Signed-off-by: Mario Limonciello <mario.limonciello@amd.com> >> --- >> drivers/acpi/platform_profile.c | 27 +++++++++++++++++++++++++++ >> 1 file changed, 27 insertions(+) >> >> diff --git a/drivers/acpi/platform_profile.c >> b/drivers/acpi/platform_profile.c >> index ef6af2c655524..4e2eda18f7f5f 100644 >> --- a/drivers/acpi/platform_profile.c >> +++ b/drivers/acpi/platform_profile.c >> @@ -25,8 +25,35 @@ static_assert(ARRAY_SIZE(profile_names) == >> PLATFORM_PROFILE_LAST); >> >> static DEFINE_IDA(platform_profile_ida); >> >> +/** >> + * name_show - Show the name of the profile handler >> + * @dev: The device >> + * @attr: The attribute >> + * @buf: The buffer to write to >> + * Return: The number of bytes written >> + */ >> +static ssize_t name_show(struct device *dev, >> + struct device_attribute *attr, >> + char *buf) >> +{ >> + struct platform_profile_handler *handler = dev_get_drvdata(dev); >> + >> + scoped_cond_guard(mutex_intr, return -ERESTARTSYS, &profile_lock) { >> + return sysfs_emit(buf, "%s\n", handler->name); >> + } >> + return -ERESTARTSYS; > > I still have a bad feeling about the locking inside the class > attributes... > > Can we assume that no sysfs accesses occur after unregistering the > class device? > > Even if this is not the case then the locking fails to protect the > platform_profile_handler here. > If the device is unregistered right after dev_get_drvdata() was > called, then we would sill operate > on possibly stale data once we take the profile_lock. > > Does someone have any clue how sysfs attributes act during removal? > I think i found the answer to my questions inside this patch series: https://lore.kernel.org/linux-kernel/1390951311-15325-1-git-send-email-tj@kernel.org It says that: kernfs / sysfs implement the "sever" semantic for userland accesses. When a node is removed, no further userland operations are allowed and the in-flight ones are drained before removal is finished. This makes policing post-mortem userland accesses trivial for its users. In this case taking the profile_lock when reading/writing class attributes seems to be unnecessary. Please remove the unnecessary locking inside the class attributes. Thanks, Armin Wolf > Thanks, > Armin Wolf > >> +} >> + >> +static DEVICE_ATTR_RO(name); >> +static struct attribute *profile_attrs[] = { >> + &dev_attr_name.attr, >> + NULL >> +}; >> +ATTRIBUTE_GROUPS(profile); >> + >> static const struct class platform_profile_class = { >> .name = "platform-profile", >> + .dev_groups = profile_groups, >> }; >> >> static ssize_t platform_profile_choices_show(struct device *dev, >
On 11/18/2024 18:28, Armin Wolf wrote: > Am 18.11.24 um 20:43 schrieb Armin Wolf: > >> Am 09.11.24 um 05:41 schrieb Mario Limonciello: >> >>> The name attribute shows the name of the associated platform profile >>> handler. >>> >>> Tested-by: Mark Pearson <mpearson-lenovo@squebb.ca> >>> Signed-off-by: Mario Limonciello <mario.limonciello@amd.com> >>> --- >>> drivers/acpi/platform_profile.c | 27 +++++++++++++++++++++++++++ >>> 1 file changed, 27 insertions(+) >>> >>> diff --git a/drivers/acpi/platform_profile.c >>> b/drivers/acpi/platform_profile.c >>> index ef6af2c655524..4e2eda18f7f5f 100644 >>> --- a/drivers/acpi/platform_profile.c >>> +++ b/drivers/acpi/platform_profile.c >>> @@ -25,8 +25,35 @@ static_assert(ARRAY_SIZE(profile_names) == >>> PLATFORM_PROFILE_LAST); >>> >>> static DEFINE_IDA(platform_profile_ida); >>> >>> +/** >>> + * name_show - Show the name of the profile handler >>> + * @dev: The device >>> + * @attr: The attribute >>> + * @buf: The buffer to write to >>> + * Return: The number of bytes written >>> + */ >>> +static ssize_t name_show(struct device *dev, >>> + struct device_attribute *attr, >>> + char *buf) >>> +{ >>> + struct platform_profile_handler *handler = dev_get_drvdata(dev); >>> + >>> + scoped_cond_guard(mutex_intr, return -ERESTARTSYS, &profile_lock) { >>> + return sysfs_emit(buf, "%s\n", handler->name); >>> + } >>> + return -ERESTARTSYS; >> >> I still have a bad feeling about the locking inside the class >> attributes... >> >> Can we assume that no sysfs accesses occur after unregistering the >> class device? >> >> Even if this is not the case then the locking fails to protect the >> platform_profile_handler here. >> If the device is unregistered right after dev_get_drvdata() was >> called, then we would sill operate >> on possibly stale data once we take the profile_lock. >> >> Does someone have any clue how sysfs attributes act during removal? >> > I think i found the answer to my questions inside this patch series: > https://lore.kernel.org/linux-kernel/1390951311-15325-1-git-send-email- > tj@kernel.org > > It says that: > > kernfs / sysfs implement the "sever" semantic for userland accesses. > When a node is removed, no further userland operations are allowed and > the in-flight ones are drained before removal is finished. This makes > policing post-mortem userland accesses trivial for its users. > > In this case taking the profile_lock when reading/writing class > attributes seems to be unnecessary. > Please remove the unnecessary locking inside the class attributes. > Before I respin a v7, let's make sure we're agreed on which things need locking and which don't. Functions that check if a lock is held: _store_class_profile() _notify_class_profile() get_class_profile() _aggregate_choices() Functions that take a lock: name_show() choices_show() profile_show() profile_store() platform_profile_choices_show() platform_profile_show() platform_profile_store() platform_profile_cycle() platform_profile_register() platform_profile_remove() Functions that don't take or check for a lock (these are intermediary and things they call check for one): _aggregate_profiles() _store_and_notify() Are you suggesting that basically these 4 can drop taking the lock? name_show() choices_show() profile_show() profile_store() I think the show() ones I can get behind, but I'm worried about profile_store(), particularly as it pertains to the other callers of _store_class_profile() because it's incongruent how the other callers would use it then. Can we perhaps just drop it for the 3 class attribute show() ones? LMK.
Am 19.11.24 um 05:09 schrieb Mario Limonciello: > On 11/18/2024 18:28, Armin Wolf wrote: >> Am 18.11.24 um 20:43 schrieb Armin Wolf: >> >>> Am 09.11.24 um 05:41 schrieb Mario Limonciello: >>> >>>> The name attribute shows the name of the associated platform profile >>>> handler. >>>> >>>> Tested-by: Mark Pearson <mpearson-lenovo@squebb.ca> >>>> Signed-off-by: Mario Limonciello <mario.limonciello@amd.com> >>>> --- >>>> drivers/acpi/platform_profile.c | 27 +++++++++++++++++++++++++++ >>>> 1 file changed, 27 insertions(+) >>>> >>>> diff --git a/drivers/acpi/platform_profile.c >>>> b/drivers/acpi/platform_profile.c >>>> index ef6af2c655524..4e2eda18f7f5f 100644 >>>> --- a/drivers/acpi/platform_profile.c >>>> +++ b/drivers/acpi/platform_profile.c >>>> @@ -25,8 +25,35 @@ static_assert(ARRAY_SIZE(profile_names) == >>>> PLATFORM_PROFILE_LAST); >>>> >>>> static DEFINE_IDA(platform_profile_ida); >>>> >>>> +/** >>>> + * name_show - Show the name of the profile handler >>>> + * @dev: The device >>>> + * @attr: The attribute >>>> + * @buf: The buffer to write to >>>> + * Return: The number of bytes written >>>> + */ >>>> +static ssize_t name_show(struct device *dev, >>>> + struct device_attribute *attr, >>>> + char *buf) >>>> +{ >>>> + struct platform_profile_handler *handler = dev_get_drvdata(dev); >>>> + >>>> + scoped_cond_guard(mutex_intr, return -ERESTARTSYS, >>>> &profile_lock) { >>>> + return sysfs_emit(buf, "%s\n", handler->name); >>>> + } >>>> + return -ERESTARTSYS; >>> >>> I still have a bad feeling about the locking inside the class >>> attributes... >>> >>> Can we assume that no sysfs accesses occur after unregistering the >>> class device? >>> >>> Even if this is not the case then the locking fails to protect the >>> platform_profile_handler here. >>> If the device is unregistered right after dev_get_drvdata() was >>> called, then we would sill operate >>> on possibly stale data once we take the profile_lock. >>> >>> Does someone have any clue how sysfs attributes act during removal? >>> >> I think i found the answer to my questions inside this patch series: >> https://lore.kernel.org/linux-kernel/1390951311-15325-1-git-send-email- >> tj@kernel.org >> >> It says that: >> >> kernfs / sysfs implement the "sever" semantic for userland >> accesses. >> When a node is removed, no further userland operations are >> allowed and >> the in-flight ones are drained before removal is finished. This >> makes >> policing post-mortem userland accesses trivial for its users. >> >> In this case taking the profile_lock when reading/writing class >> attributes seems to be unnecessary. >> Please remove the unnecessary locking inside the class attributes. >> > > Before I respin a v7, let's make sure we're agreed on which things > need locking and which don't. > > Functions that check if a lock is held: > _store_class_profile() > _notify_class_profile() > get_class_profile() > _aggregate_choices() > > Functions that take a lock: > name_show() > choices_show() > profile_show() > profile_store() > platform_profile_choices_show() > platform_profile_show() > platform_profile_store() > platform_profile_cycle() > platform_profile_register() > platform_profile_remove() > > Functions that don't take or check for a lock (these are intermediary > and things they call check for one): > _aggregate_profiles() > _store_and_notify() > > Are you suggesting that basically these 4 can drop taking the lock? > name_show() > choices_show() > profile_show() > profile_store() > > I think the show() ones I can get behind, but I'm worried about > profile_store(), particularly as it pertains to the other callers of > _store_class_profile() because it's incongruent how the other callers > would use it then. > > Can we perhaps just drop it for the 3 class attribute show() ones? I think so, i also remembered that profile_store() needs to keep taking the lock in case platform_profile_cycle() is currently running. Can you also remove the second call to dev_get_drvdata() in _store_class_profile()? Thanks, Armin Wolf > > LMK. > > >
On 11/19/2024 06:26, Armin Wolf wrote: > Am 19.11.24 um 05:09 schrieb Mario Limonciello: > >> On 11/18/2024 18:28, Armin Wolf wrote: >>> Am 18.11.24 um 20:43 schrieb Armin Wolf: >>> >>>> Am 09.11.24 um 05:41 schrieb Mario Limonciello: >>>> >>>>> The name attribute shows the name of the associated platform profile >>>>> handler. >>>>> >>>>> Tested-by: Mark Pearson <mpearson-lenovo@squebb.ca> >>>>> Signed-off-by: Mario Limonciello <mario.limonciello@amd.com> >>>>> --- >>>>> drivers/acpi/platform_profile.c | 27 +++++++++++++++++++++++++++ >>>>> 1 file changed, 27 insertions(+) >>>>> >>>>> diff --git a/drivers/acpi/platform_profile.c >>>>> b/drivers/acpi/platform_profile.c >>>>> index ef6af2c655524..4e2eda18f7f5f 100644 >>>>> --- a/drivers/acpi/platform_profile.c >>>>> +++ b/drivers/acpi/platform_profile.c >>>>> @@ -25,8 +25,35 @@ static_assert(ARRAY_SIZE(profile_names) == >>>>> PLATFORM_PROFILE_LAST); >>>>> >>>>> static DEFINE_IDA(platform_profile_ida); >>>>> >>>>> +/** >>>>> + * name_show - Show the name of the profile handler >>>>> + * @dev: The device >>>>> + * @attr: The attribute >>>>> + * @buf: The buffer to write to >>>>> + * Return: The number of bytes written >>>>> + */ >>>>> +static ssize_t name_show(struct device *dev, >>>>> + struct device_attribute *attr, >>>>> + char *buf) >>>>> +{ >>>>> + struct platform_profile_handler *handler = dev_get_drvdata(dev); >>>>> + >>>>> + scoped_cond_guard(mutex_intr, return -ERESTARTSYS, >>>>> &profile_lock) { >>>>> + return sysfs_emit(buf, "%s\n", handler->name); >>>>> + } >>>>> + return -ERESTARTSYS; >>>> >>>> I still have a bad feeling about the locking inside the class >>>> attributes... >>>> >>>> Can we assume that no sysfs accesses occur after unregistering the >>>> class device? >>>> >>>> Even if this is not the case then the locking fails to protect the >>>> platform_profile_handler here. >>>> If the device is unregistered right after dev_get_drvdata() was >>>> called, then we would sill operate >>>> on possibly stale data once we take the profile_lock. >>>> >>>> Does someone have any clue how sysfs attributes act during removal? >>>> >>> I think i found the answer to my questions inside this patch series: >>> https://lore.kernel.org/linux-kernel/1390951311-15325-1-git-send-email- >>> tj@kernel.org >>> >>> It says that: >>> >>> kernfs / sysfs implement the "sever" semantic for userland >>> accesses. >>> When a node is removed, no further userland operations are >>> allowed and >>> the in-flight ones are drained before removal is finished. This >>> makes >>> policing post-mortem userland accesses trivial for its users. >>> >>> In this case taking the profile_lock when reading/writing class >>> attributes seems to be unnecessary. >>> Please remove the unnecessary locking inside the class attributes. >>> >> >> Before I respin a v7, let's make sure we're agreed on which things >> need locking and which don't. >> >> Functions that check if a lock is held: >> _store_class_profile() >> _notify_class_profile() >> get_class_profile() >> _aggregate_choices() >> >> Functions that take a lock: >> name_show() >> choices_show() >> profile_show() >> profile_store() >> platform_profile_choices_show() >> platform_profile_show() >> platform_profile_store() >> platform_profile_cycle() >> platform_profile_register() >> platform_profile_remove() >> >> Functions that don't take or check for a lock (these are intermediary >> and things they call check for one): >> _aggregate_profiles() >> _store_and_notify() >> >> Are you suggesting that basically these 4 can drop taking the lock? >> name_show() >> choices_show() >> profile_show() >> profile_store() >> >> I think the show() ones I can get behind, but I'm worried about >> profile_store(), particularly as it pertains to the other callers of >> _store_class_profile() because it's incongruent how the other callers >> would use it then. >> >> Can we perhaps just drop it for the 3 class attribute show() ones? > > I think so, i also remembered that profile_store() needs to keep taking > the lock in case platform_profile_cycle() is currently > running. Actually considering this, we need to keep it on profile_show() too then for the exact same reason. I will drop it for choices and name though. > > Can you also remove the second call to dev_get_drvdata() in > _store_class_profile()? > Sure. > Thanks, > Armin Wolf > >> >> LMK. >> >> >>
diff --git a/drivers/acpi/platform_profile.c b/drivers/acpi/platform_profile.c index ef6af2c655524..4e2eda18f7f5f 100644 --- a/drivers/acpi/platform_profile.c +++ b/drivers/acpi/platform_profile.c @@ -25,8 +25,35 @@ static_assert(ARRAY_SIZE(profile_names) == PLATFORM_PROFILE_LAST); static DEFINE_IDA(platform_profile_ida); +/** + * name_show - Show the name of the profile handler + * @dev: The device + * @attr: The attribute + * @buf: The buffer to write to + * Return: The number of bytes written + */ +static ssize_t name_show(struct device *dev, + struct device_attribute *attr, + char *buf) +{ + struct platform_profile_handler *handler = dev_get_drvdata(dev); + + scoped_cond_guard(mutex_intr, return -ERESTARTSYS, &profile_lock) { + return sysfs_emit(buf, "%s\n", handler->name); + } + return -ERESTARTSYS; +} + +static DEVICE_ATTR_RO(name); +static struct attribute *profile_attrs[] = { + &dev_attr_name.attr, + NULL +}; +ATTRIBUTE_GROUPS(profile); + static const struct class platform_profile_class = { .name = "platform-profile", + .dev_groups = profile_groups, }; static ssize_t platform_profile_choices_show(struct device *dev,