Skip to contentSkip to conversation

Generic interface for accelerometers (AMS, HDAPS, ...)

DISCUSSION Henrique de Moraes Holschuh 9 participants 15 messages Jul 3, 2006 – Jul 6, 2006

lore
Share & tools
lore
#0 lore raw

We have now (or we are about to have, anyway) two drivers that export accelerometer data: IBM's HDAPS, and Apple's AMS. More accelerometer drivers could be coming in the future.

Both drivers export one common set of data (accelerometer output), and some extra information that is not related to acellerometers. Both have at least two functionality goals in common: to export accelerometer data to userspace, and also to allow somehow for HD head parking when the accelerometer detects a potentially incoming impact.

User?space utilities that interface to accelerometers like hdapsd and smackpad could benefit from a common interface, so that they work with HDAPS, AMS, and any other future accelerometer drivers.

If any kernel-space functionality needed by HDAPS and AMS is also shared, we would have benefits there too. Some examples I can think of are: generic HD queue-freeze and HD head parking, and a generic acellerometer-driven joystick event interface. This would also enable generic userspace notifiers that the HD heads have been parked, etc.

I am posting this message as a head's up for the two projects (HDAPS, AMS) to make sure that they are actively aware of each other. I do so in hope that we can work in a joint, generic interface and port both drivers to this new interface in the near future, and also that we can make as much functionality shared between the two drivers as possible.

Signature · 4 lines
  "One disk to rule them all, One disk to find them. One disk to bring
  them all and in the darkness grind them. In the Land of Redmond
  where the shadows lie." -- The Silicon Valley Tarot
  Henrique Holschuh
#1 lore raw
Quoted 20 lines · Henrique de Moraes Holschuh

We have now (or we are about to have, anyway) two drivers that export accelerometer data: IBM's HDAPS, and Apple's AMS. More accelerometer drivers could be coming in the future.

Both drivers export one common set of data (accelerometer output), and some extra information that is not related to acellerometers. Both have at least two functionality goals in common: to export accelerometer data to userspace, and also to allow somehow for HD head parking when the accelerometer detects a potentially incoming impact.

User?space utilities that interface to accelerometers like hdapsd and smackpad could benefit from a common interface, so that they work with HDAPS, AMS, and any other future accelerometer drivers.

If any kernel-space functionality needed by HDAPS and AMS is also shared, we would have benefits there too. Some examples I can think of are: generic HD queue-freeze and HD head parking, and a generic acellerometer-driven joystick event interface. This would also enable generic userspace notifiers that the HD heads have been parked, etc.

Quoted 5 lines

I am posting this message as a head's up for the two projects (HDAPS, AMS) to make sure that they are actively aware of each other. I do so in hope that we can work in a joint, generic interface and port both drivers to this new interface in the near future, and also that we can make as much functionality shared between the two drivers as possible.

Just use input infrastructure and be done with that? You can do parking from userspace.

Only piece of puzzle missing is marking that input device as "this accelerometer actually watches the whole device".
Pavel

Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
#2 lore raw
Quoted 2 lines · Pavel Machek

Just use input infrastructure and be done with that? You can do parking from userspace.

Will moving the hdapsd userspace daemon from sysfs polling to the input infrastructure cause a noticable latency increase compared to polling sysfs? This functionality is highly time-critical.

Also, there's a small issue with polling frequency. hdapsd needs a fairly high frequency (say, 50Hz) to gather statistics and keep response latency low, whereas the hdaps driver's internal polling (routing to the input infrastructure) is currently done at only 20Hz. We'll need to increase the latter, thereby slightly increasing system load when hdaps isn't running.

In terms of magnitude, an hdaps readout takes ~70usec on average and (very rarely) up to ~1msec. This is with the tp_smapi [1] patches to hdaps, which add the checks and retries needed to ensure the readout is valid.

BTW, can the driver tell when nothing is accessing its input device, and avoid polling in that case?

Shem

[1] http://thinkwiki.org/wiki/tp_smapi

#3 lore raw

[lm-sensors] [Hdaps-devel] Generic interface for accelerometers

Quoted 2 lines · Henrique de Moraes Holschuh

If any kernel-space functionality needed by HDAPS and AMS is also shared, we would have benefits there too.

[snip]

Quoted 1 line

queue-freeze and HD head parking

Is there any difference between the freeze+park functionality needed by the different laptops? The hdaps_protect patch has a well thought out generic interface, and can probably serve all projects. What's the prospect of seeing it in mainline? (Of course, the recently reported hangs need to be addressed first.)

Shem

#4 lore raw
In reply to Shem Multinymous

Hi!

Quoted 6 lines
Quoted 2 lines

Just use input infrastructure and be done with that? You can do parking from userspace.

Will moving the hdapsd userspace daemon from sysfs polling to the input infrastructure cause a noticable latency increase compared to polling sysfs? This functionality is highly time-critical.

Well, for input you should need no polling. input tells you when data is ready.

Quoted 2 lines

BTW, can the driver tell when nothing is accessing its input device, and avoid polling in that case?

Ask vojtech/dmitry, I guess.
Pavel

Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
#5 lore raw
In reply to Shem Multinymous

On July 04 at 06:26 EDT, Shem Multinymous hastily scribbled:

Quoted 10 lines

Will moving the hdapsd userspace daemon from sysfs polling to the input infrastructure cause a noticable latency increase compared to polling sysfs? This functionality is highly time-critical.

Also, there's a small issue with polling frequency. hdapsd needs a fairly high frequency (say, 50Hz) to gather statistics and keep response latency low, whereas the hdaps driver's internal polling (routing to the input infrastructure) is currently done at only 20Hz. We'll need to increase the latter, thereby slightly increasing system load when hdaps isn't running.

Just out of curiousity, is there any reason that these hard drive parking schemes *aren't* implemented entirely in the kernel? Wouldn't implementing it in the kernel give it much lower latency?

Signature · 4 lines
Thomas Tuttle ([email protected])
Get Thunderbird: Reclaim your inbox. mozilla.org/products/thunderbird
aim/y!m:thinkinginbinary; icq:198113263; jabber:[email protected]
msn: [email protected]; pgp: 0xAF5112C6
#6 lore raw
In reply to Shem Multinymous
Quoted 4 lines · Shem Multinymous

BTW, can the driver tell when nothing is accessing its input device, and avoid polling in that case?

Yes (->open is called when there is a client opening one of the input interfaces), but I don't think that "oh shit I dropped my laptop" events belong to input layer.

Signature · 1 line
Dmitry
#7 lore raw
In reply to Shem Multinymous

[I hope I managed to not break the thread]

Quoted 6 lines

Also, there's a small issue with polling frequency. hdapsd needs a fairly high frequency (say, 50Hz) to gather statistics and keep response latency low, whereas the hdaps driver's internal polling (routing to the input infrastructure) is currently done at only 20Hz. We'll need to increase the latter, thereby slightly increasing system load when hdaps isn't running.

Note that with AMS we're better off -- it has two interrupts telling us when something is wrong.

Hence, most of the discussion about loads of input values only applies to hdaps, the actual head-park functionality can be implemented with AMS without ever reading any sensor values.

Hence we also need much less complexity in userland -- once an interrupt comes in we trigger the hd park...

johannes
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 828 bytes
Desc: This is a digitally signed message part
Url : http://lists.lm-sensors.org/pipermail/lm-sensors/attachments/20060705/d69ff9ed/attachment-0001.bin

#8 lore raw
Quoted 10 lines · Johannes Berg
Quoted 6 lines

Also, there's a small issue with polling frequency. hdapsd needs a fairly high frequency (say, 50Hz) to gather statistics and keep response latency low, whereas the hdaps driver's internal polling (routing to the input infrastructure) is currently done at only 20Hz. We'll need to increase the latter, thereby slightly increasing system load when hdaps isn't running.

Note that with AMS we're better off -- it has two interrupts telling us when something is wrong.

Quoted 6 lines

Hence, most of the discussion about loads of input values only applies to hdaps, the actual head-park functionality can be implemented with AMS without ever reading any sensor values.

Hence we also need much less complexity in userland -- once an interrupt comes in we trigger the hd park...

Looks nice. For AMS, then, the userspace daemon can choose between deciding for itself when to park heads using accel data, or to trust the firmware and do it when told, or even to do both.

IBM *could* have done the same, since they already have an H8 microcontroller looking at that accelerometer :( Anyway, IMHO it would be good to have AMS-like behaviour where the kernel driver exposes head-park events to userspace in the generic interface.

Signature · 4 lines
  "One disk to rule them all, One disk to find them. One disk to bring
  them all and in the darkness grind them. In the Land of Redmond
  where the shadows lie." -- The Silicon Valley Tarot
  Henrique Holschuh
#9 lore raw
In reply to Pavel Machek
Quoted 2 lines · Pavel Machek

Just use input infrastructure and be done with that? You can do parking from userspace.

The command to execute the freeze (and for how long) can certaily come from userspace, and the logic behind that command can be an userspace high-priority task, yes (that's how it is done in HDAPS anyway). However, to implement the disk head unload you need kernel code (you need to freeze the IO queues for a while too, not just unload the heads).

Anyway, the input infrastructure is good to emulate a mouse/joystick, but we probably want to support the following generic data channels:

1. Absolute acceleration output stream (either in hardware units, or normalized to G or mG if at all possible) for X, Y, Z. Does the input infrastructure work well for reporting absolute numbers in a reasonably constant rate?

This is what HD head unload policy daemons want to know, and the stream usually needs to offer datapoints somewhere between 20Hz and 100Hz without excessive jitter to be useful for more advanced filtering (like masking-off periodic movement such as the one caused by train tracks).

2. As easy-to-use as possible joystick and mouse emulation (for toys and gaming), which probably means auto-calibrating ones when possible and thus making them unsuitable channels for (1) above.

3. Control of accelerometer parameters:
3a. Report of accelerometer type (hdaps, ams, etc) and other metadata
(name, location, what it is measuring (system accel, hd-bay accel...))
3b. Accelerometer polling frequency
3c. Enable/disable joystick and mouse emulation control etc.

And, of course, userspace must be able to easily find the accelerometer outputs and diferentiate them from other joystick/mouse interfaces. This is a task for udev, I suppose.

It is also very helpful to be able to know when something is using any of the accelerometer output channels, so as to shut it down (or stop talking to it) when it is uneeded. I don't know about AMS, but talking to HDAPS when you don't need to does waste enough system resources and power to actually justify implementing this.

If we cannot detect automatically when userspace is paying attention to the accelerometer through the input layer, then that means we will need the enable/disable functionality already provided by the device model through sysfs.

Quoted 2 lines

Only piece of puzzle missing is marking that input device as "this accelerometer actually watches the whole device".

Well, it is very important to be able to easily find all data and channels pertaining for a given accelerometer, and the accelerometer metadata should indeed give userspace an idea of WHAT it is measuring the acceleration of :-)

Signature · 4 lines
  "One disk to rule them all, One disk to find them. One disk to bring
  them all and in the darkness grind them. In the Land of Redmond
  where the shadows lie." -- The Silicon Valley Tarot
  Henrique Holschuh
#10 lore raw

Hi!

Quoted 8 lines
Quoted 2 lines

Just use input infrastructure and be done with that? You can do parking from userspace.

The command to execute the freeze (and for how long) can certaily come from userspace, and the logic behind that command can be an userspace high-priority task, yes (that's how it is done in HDAPS anyway). However, to implement the disk head unload you need kernel code (you need to freeze the IO queues for a while too, not just unload the heads).

Okay, yes, this part is needed. It is useful for more than accelerometers...

Quoted 8 lines

Anyway, the input infrastructure is good to emulate a mouse/joystick, but we probably want to support the following generic data channels:

1. Absolute acceleration output stream (either in hardware units, or normalized to G or mG if at all possible) for X, Y, Z. Does the input infrastructure work well for reporting absolute numbers in a reasonably constant rate?

Quoted 8 lines

This is what HD head unload policy daemons want to know, and the stream usually needs to offer datapoints somewhere between 20Hz and 100Hz without excessive jitter to be useful for more advanced filtering (like masking-off periodic movement such as the one caused by train tracks).

2. As easy-to-use as possible joystick and mouse emulation (for toys and gaming), which probably means auto-calibrating ones when possible and thus making them unsuitable channels for (1) above.

Well, I'd just provide 1. Providing 2 seems like task for some userspace filtering library.

Quoted 4 lines

3. Control of accelerometer parameters:
3a. Report of accelerometer type (hdaps, ams, etc) and other metadata
(name, location, what it is measuring (system accel, hd-bay accel...))

Well, if your system is accelerating at different speed than hd-bay, then your machine is either _way_ too big, or you are in bad trouble.

Ask Dmitry or vojtech, I believe input can provide such metadata. (I left most of the quoted text so that vojtech can reply easily).

Pavel

Quoted 27 lines

3b. Accelerometer polling frequency
3c. Enable/disable joystick and mouse emulation control etc.

And, of course, userspace must be able to easily find the accelerometer outputs and diferentiate them from other joystick/mouse interfaces. This is a task for udev, I suppose.

It is also very helpful to be able to know when something is using any of the accelerometer output channels, so as to shut it down (or stop talking to it) when it is uneeded. I don't know about AMS, but talking to HDAPS when you don't need to does waste enough system resources and power to actually justify implementing this.

If we cannot detect automatically when userspace is paying attention to the accelerometer through the input layer, then that means we will need the enable/disable functionality already provided by the device model through sysfs.

Quoted 2 lines

Only piece of puzzle missing is marking that input device as "this accelerometer actually watches the whole device".

Well, it is very important to be able to easily find all data and channels pertaining for a given accelerometer, and the accelerometer metadata should indeed give userspace an idea of WHAT it is measuring the acceleration of :-)

Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
#11 lore raw
Quoted 8 lines · Pavel Machek
Quoted 8 lines

Anyway, the input infrastructure is good to emulate a mouse/joystick, but we probably want to support the following generic data channels:

1. Absolute acceleration output stream (either in hardware units, or normalized to G or mG if at all possible) for X, Y, Z. Does the input infrastructure work well for reporting absolute numbers in a reasonably constant rate?

Quoted 4 lines
Quoted 4 lines

This is what HD head unload policy daemons want to know, and the stream usually needs to offer datapoints somewhere between 20Hz and 100Hz without excessive jitter to be useful for more advanced filtering (like masking-off periodic movement such as the one caused by train tracks).

Yes, this is what the input subsystem normally works with - 32-bit absolute (and relative) numbers at rates around 100 Hz. It's how the data from joysticks (and mice) look like.

Quoted 6 lines
Quoted 3 lines

2. As easy-to-use as possible joystick and mouse emulation (for toys and gaming), which probably means auto-calibrating ones when possible and thus making them unsuitable channels for (1) above.

Well, I'd just provide 1. Providing 2 seems like task for some userspace filtering library.

Assuming you use the ABS_X, ABS_Y, and ABS_Z for the acceleration (which is not necessarily ideal), you'll get a joystick emulation automagically through the input subsystem without any additional coding.

Quoted 10 lines
Quoted 4 lines

3. Control of accelerometer parameters:
3a. Report of accelerometer type (hdaps, ams, etc) and other metadata
(name, location, what it is measuring (system accel, hd-bay accel...))

Well, if your system is accelerating at different speed than hd-bay, then your machine is either _way_ too big, or you are in bad trouble.

Ask Dmitry or vojtech, I believe input can provide such metadata. (I left most of the quoted text so that vojtech can reply easily).

It currently doesn't. It reports the device ID as a couple of 16-bit integers, and how the input subsystem sees it. There is currently no way to get at device-specific data, although this could be reasonably easily done through sysfs I believe.

Quoted 1 line
Quoted 1 line

3b. Accelerometer polling frequency

Module parameter, changeable via sysfs?

Quoted 6 lines
Quoted 6 lines

3c. Enable/disable joystick and mouse emulation control etc.

And, of course, userspace must be able to easily find the accelerometer outputs and diferentiate them from other joystick/mouse interfaces. This is a task for udev, I suppose.

Yup.

Quoted 3 lines
Quoted 3 lines

It is also very helpful to be able to know when something is using any of the accelerometer output channels, so as to shut it down (or stop talking to it) when it is uneeded.

This is done automatically by the input subsystem. When the last user close()s any of the related devices, the driver gets notified that it can stop the hardware.

Quoted 3 lines
Quoted 3 lines

I don't know about AMS, but talking to HDAPS when
you don't need to does waste enough system resources and power to actually justify implementing this.

I'd doubt any of the accelerometer implementations would consume much power or CPU.

Quoted 11 lines
Quoted 11 lines

If we cannot detect automatically when userspace is paying attention to the accelerometer through the input layer, then that means we will need the enable/disable functionality already provided by the device model through sysfs.

Quoted 2 lines

Only piece of puzzle missing is marking that input device as "this accelerometer actually watches the whole device".

Well, it is very important to be able to easily find all data and channels pertaining for a given accelerometer, and the accelerometer metadata should indeed give userspace an idea of WHAT it is measuring the acceleration of

Signature · 2 lines
Vojtech Pavlik
Director SuSE Labs
#12 lore raw
Quoted 6 lines · Vojtech Pavlik
Quoted 3 lines
Quoted 3 lines

I don't know about AMS, but talking to HDAPS when
you don't need to does waste enough system resources and power to actually justify implementing this.

I'd doubt any of the accelerometer implementations would consume much power or CPU.

It is not just CPU or power, although IMO striving for perfection on power management in a laptop is almost always a good thing if it can be done safely.

HDAPS talks to the embedded controller using IO over the LPC bus, and not to the accelerometer chip or to a simple A/D i2c chip which is used excusively for accelerometer access. The EC interface for HDAPS data retrieval is not friendly to any errors, and hardlocks the machine somehow if any firmware bugs hit or if we violate any of the rules (that are not written anywhere) about how to access the EC without geting the SMBIOS unhappy.

So, turning off HDAPS polling while it is not necessary really looks like a good idea overall.

We are investigating the ACPI global lock as a way to at least get the SMBIOS to stay away from the EC while we talk to it, but we don't know if the entire SMBIOS firmware respects that lock.

Signature · 4 lines
  "One disk to rule them all, One disk to find them. One disk to bring
  them all and in the darkness grind them. In the Land of Redmond
  where the shadows lie." -- The Silicon Valley Tarot
  Henrique Holschuh
#13 lore raw

Hi!

Quoted 6 lines
Quoted 6 lines
Quoted 3 lines
Quoted 3 lines

I don't know about AMS, but talking to HDAPS when
you don't need to does waste enough system resources and power to actually justify implementing this.

I'd doubt any of the accelerometer implementations would consume much power or CPU.

Actually polling at 100Hz probably _is_ going to be measurable on modern CPUs -- because it means that CPU has to exit C4. It will not be too bad, hopefully.

Pavel

Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
#14 lore raw
In reply to Pavel Machek

(Does this conversation really need to happen on 3 different mailing lists at once?)

Quoted 7 lines
Quoted 4 lines

3. Control of accelerometer parameters:
3a. Report of accelerometer type (hdaps, ams, etc) and other metadata
(name, location, what it is measuring (system accel, hd-bay accel...))

Well, if your system is accelerating at different speed than hd-bay, then your machine is either _way_ too big, or you are in bad trouble.

Not necessarily. If the system is spinning, each point has its own acceleration vector. Having two accelerometers in a system would be a convenient way to detect that kind of movement, and I wouldn't be surprised to see it happen.

Signature · 1 line
Jean Delvare