Message ID | 20211220174155.40239-1-cristian.marussi@arm.com (mailing list archive) |
---|---|
Headers | show |
Series | Sensor readings fixes | expand |
On Mon, Dec 20, 2021 at 05:41:53PM +0000, Cristian Marussi wrote: > Hi, > > this was supposed to be an easy fix on how sensor readings are handled > across different FW versions while maintaining backward compatibility, > but the solution raised for me more questions than the issue itself... > ...so I posted as an RFC. > > In a nutshell, since SCMI FWv3.0 spec, sensors SCMI_READING_GET command > can report axis and timestamps too, beside readings, so a brand new > scmi_reading_get_timestamped protocol operation was exposed (used by IIO) > while the old scmi_reading_get was kept as it was, already used by HWMON > subsystem for other classes of sensors. > > Unfortunately, also the flavour of reported values changed from unsigned > to signed with v3.0, so if you end-up on a system running against an SCMI > v3.0 FW platform you could end up reading a negative value and interpreting > it as a big positive since scmi_reading_get reports only u64. > > 01/02 simply takes care, when a FW >= 3.0 is detected, to return an error > to any scmi_reading_get request if that would result in tryinh to carry > a negative value into an u64. > > So this should rectify the API exposed by SCMI sensor and make it > consistent in general, in such a way that a user calling it won't risk to > receive a false big-positive which was indeed a 2-complement negative from > the perpective of the SCMI fw. > > So far so good...sort of...since, to make things more dire, the HWMON > interface, which is the only current upstream user of scmi_reading_get > DOES allow indeed to report to the HWMON core negative values, so it was > just that we were silently interpreting u64 as s64 :P ... > > ...as a consequence the fix above to the SCMI API will potentially break > this undocumented behaviour of our only scmi_reading_get user. > > Additionally, while looking at this, I realized that for similar reasons > even on systems running the current SCMI stack API and an old FW <=2.0 > the current HWMON read is potentially broken, since when the FW reports > a very big and real positive number we'll report it as a signed long to > the HWMON core, so turning it wrongly into a negative report: for this > reason 02/02 adds a check inside scmi-hwmon to filter out, reporting > errors, any result reported by scmi_reading_get so big as to be considered > a negative in 2-complement... > > ...and this will probably break even more the undocumented behaviours... > > Any feedback welcome ! Hi, any feedback on this ? (...before I forgot again :D) Thanks, Cristian
On 3/25/22 04:41, Cristian Marussi wrote: > On Mon, Dec 20, 2021 at 05:41:53PM +0000, Cristian Marussi wrote: >> Hi, >> >> this was supposed to be an easy fix on how sensor readings are handled >> across different FW versions while maintaining backward compatibility, >> but the solution raised for me more questions than the issue itself... >> ...so I posted as an RFC. >> >> In a nutshell, since SCMI FWv3.0 spec, sensors SCMI_READING_GET command >> can report axis and timestamps too, beside readings, so a brand new >> scmi_reading_get_timestamped protocol operation was exposed (used by IIO) >> while the old scmi_reading_get was kept as it was, already used by HWMON >> subsystem for other classes of sensors. >> >> Unfortunately, also the flavour of reported values changed from unsigned >> to signed with v3.0, so if you end-up on a system running against an SCMI >> v3.0 FW platform you could end up reading a negative value and interpreting >> it as a big positive since scmi_reading_get reports only u64. >> >> 01/02 simply takes care, when a FW >= 3.0 is detected, to return an error >> to any scmi_reading_get request if that would result in tryinh to carry >> a negative value into an u64. >> >> So this should rectify the API exposed by SCMI sensor and make it >> consistent in general, in such a way that a user calling it won't risk to >> receive a false big-positive which was indeed a 2-complement negative from >> the perpective of the SCMI fw. >> >> So far so good...sort of...since, to make things more dire, the HWMON >> interface, which is the only current upstream user of scmi_reading_get >> DOES allow indeed to report to the HWMON core negative values, so it was >> just that we were silently interpreting u64 as s64 :P ... >> >> ...as a consequence the fix above to the SCMI API will potentially break >> this undocumented behaviour of our only scmi_reading_get user. >> >> Additionally, while looking at this, I realized that for similar reasons >> even on systems running the current SCMI stack API and an old FW <=2.0 >> the current HWMON read is potentially broken, since when the FW reports >> a very big and real positive number we'll report it as a signed long to >> the HWMON core, so turning it wrongly into a negative report: for this >> reason 02/02 adds a check inside scmi-hwmon to filter out, reporting >> errors, any result reported by scmi_reading_get so big as to be considered >> a negative in 2-complement... >> >> ...and this will probably break even more the undocumented behaviours... >> >> Any feedback welcome ! > > Hi, > > any feedback on this ? (...before I forgot again :D) Sorry for the lag, I threw these into a build and the first thing that popped is the following warning on a 32-bit ARM build: In file included from ./include/linux/bits.h:6, from ./include/linux/bitops.h:6, from ./include/linux/hwmon.h:15, from drivers/hwmon/scmi-hwmon.c:9: drivers/hwmon/scmi-hwmon.c: In function 'scmi_hwmon_read': ./include/vdso/bits.h:7:26: warning: left shift count >= width of type [-Wshift-count-overflow] #define BIT(nr) (UL(1) << (nr)) ^~ drivers/hwmon/scmi-hwmon.c:88:14: note: in expansion of macro 'BIT' if (value & BIT(63)) { ^~~ Now, in terms of functional testing it did seems to work as intended for 32-bit kernels not for 64-bit kernels where I got: # sensors scmi_sensors-virtual-0 Adapter: Virtual device [ 16.413590] hwmon hwmon0: Reported unsigned value too big. ERROR: Can't get value of subfeature temp1_input: I/O error avs_pvt_temp: N/A pmic_die_temp: +53.4 C whereas 32-bit would return the following: # sensors scmi_sensors-virtual-0 Adapter: Virtual device avs_pvt_temp: -6.7 C pmic_die_temp: +52.3 C The firmware is version 1: [ 0.044969] arm-scmi brcm_scmi@0: SCMI Protocol v1.0 'brcm-scmi:' Firmware version 0x1 I will need to revisit the specification to make sure I implemented it correctly the first time in our version of TF-A.
On Fri, Mar 25, 2022 at 02:38:06PM -0700, Florian Fainelli wrote: > On 3/25/22 04:41, Cristian Marussi wrote: > > On Mon, Dec 20, 2021 at 05:41:53PM +0000, Cristian Marussi wrote: > > > Hi, > > > > > > this was supposed to be an easy fix on how sensor readings are handled > > > across different FW versions while maintaining backward compatibility, > > > but the solution raised for me more questions than the issue itself... > > > ...so I posted as an RFC. > > > > > > In a nutshell, since SCMI FWv3.0 spec, sensors SCMI_READING_GET command > > > can report axis and timestamps too, beside readings, so a brand new > > > scmi_reading_get_timestamped protocol operation was exposed (used by IIO) > > > while the old scmi_reading_get was kept as it was, already used by HWMON > > > subsystem for other classes of sensors. > > > > > > Unfortunately, also the flavour of reported values changed from unsigned > > > to signed with v3.0, so if you end-up on a system running against an SCMI > > > v3.0 FW platform you could end up reading a negative value and interpreting > > > it as a big positive since scmi_reading_get reports only u64. > > > > > > 01/02 simply takes care, when a FW >= 3.0 is detected, to return an error > > > to any scmi_reading_get request if that would result in tryinh to carry > > > a negative value into an u64. > > > > > > So this should rectify the API exposed by SCMI sensor and make it > > > consistent in general, in such a way that a user calling it won't risk to > > > receive a false big-positive which was indeed a 2-complement negative from > > > the perpective of the SCMI fw. > > > > > > So far so good...sort of...since, to make things more dire, the HWMON > > > interface, which is the only current upstream user of scmi_reading_get > > > DOES allow indeed to report to the HWMON core negative values, so it was > > > just that we were silently interpreting u64 as s64 :P ... > > > > > > ...as a consequence the fix above to the SCMI API will potentially break > > > this undocumented behaviour of our only scmi_reading_get user. > > > > > > Additionally, while looking at this, I realized that for similar reasons > > > even on systems running the current SCMI stack API and an old FW <=2.0 > > > the current HWMON read is potentially broken, since when the FW reports > > > a very big and real positive number we'll report it as a signed long to > > > the HWMON core, so turning it wrongly into a negative report: for this > > > reason 02/02 adds a check inside scmi-hwmon to filter out, reporting > > > errors, any result reported by scmi_reading_get so big as to be considered > > > a negative in 2-complement... > > > > > > ...and this will probably break even more the undocumented behaviours... > > > > > > Any feedback welcome ! > > > > Hi, > > > > any feedback on this ? (...before I forgot again :D) > > Sorry for the lag, I threw these into a build and the first thing that > popped is the following warning on a 32-bit ARM build: > Hi Florian, thanks for the feedback first of all... > In file included from ./include/linux/bits.h:6, > from ./include/linux/bitops.h:6, > from ./include/linux/hwmon.h:15, > from drivers/hwmon/scmi-hwmon.c:9: > drivers/hwmon/scmi-hwmon.c: In function 'scmi_hwmon_read': > ./include/vdso/bits.h:7:26: warning: left shift count >= width of type > [-Wshift-count-overflow] > #define BIT(nr) (UL(1) << (nr)) > ^~ > drivers/hwmon/scmi-hwmon.c:88:14: note: in expansion of macro 'BIT' > if (value & BIT(63)) { > ^~~ > ..and sorry that the series does not seem in good shape... > Now, in terms of functional testing it did seems to work as intended for > 32-bit kernels not for 64-bit kernels where I got: > > # sensors > scmi_sensors-virtual-0 > Adapter: Virtual device > [ 16.413590] hwmon hwmon0: Reported unsigned value too big. > ERROR: Can't get value of subfeature temp1_input: I/O error > avs_pvt_temp: N/A > pmic_die_temp: +53.4 C > So this is my patch apparently breaking things....which was what I wanted to verify really :P ... the thing is that up till SCMI v2.0 (Sensor Vers <= 0x10000) the SENSOR_READING_GET command returned a single u32 reading-value by the spec, after that, starting with SCMIv3.0, the SENSOR_READING_GET returned also a timestamp and per-axis reading-values BUT these readings are now signed s32 ! So I kept the old SCMI interface as it was (used by HWMON): int (*reading_get)(const struct scmi_protocol_handle *ph, u32 sensor_id, u64 *value); and introduced a new one for timestamped/per-axis values provided by newer FW (used by SCMI IIO driver): int (*reading_get_timestamped)(const struct scmi_protocol_handle *ph, u32 sensor_id, u8 count, struct scmi_sensor_reading *readings); (which conveys timestamps and s32 values inside the *readings) The old interface pass back unsigned values only, in theory, BUT its only user HWMON hanldes also negatives, so, it sort of makes sense if the FW conveyed signed values inside an unsigned variable in the context of HWMON, (breaking the spec) since it cannot convey negatives in any other way... The whole point of this (broken) series was to try to see if I could sort of sanitizing the results depending on the backend FW version detected while maintaining backward compatibility...but the current approach of this series is deadly broken (as you had seen :<) since makes impossible really at the end for the FW to convey negative values as a whole... I'll have a thought about it but I think I'll drop this series as it is... > whereas 32-bit would return the following: > > # sensors > scmi_sensors-virtual-0 > Adapter: Virtual device > avs_pvt_temp: -6.7 C > pmic_die_temp: +52.3 C > > The firmware is version 1: > > [ 0.044969] arm-scmi brcm_scmi@0: SCMI Protocol v1.0 'brcm-scmi:' > Firmware version 0x1 > I think this works because my patch is flaky :P ... but as said I'll drop this as it is now. Thanks and sorry for the noise, Cristian