User:CesarB/cpufreq

From Openmoko

< User:CesarB
Revision as of 00:00, 27 January 2008 by CesarB (Talk | contribs)

Jump to: navigation, search

The cpufreq subsystem is a Linux kernel subsystem responsible for managing the CPU frequency. It's commonly used in laptops to reduce power usage when idle.

On the S3C2410, adjusting the CPU frequency changes the clocks for almost all devices on the chip. This means that, to write a cpufreq driver for the S3C2410, you have to also adjust some values on almost all devices.

This project aims to implement the cpufreq driver (which manages the CPU frequency transitions) and a cpufreq notifier for all the affected drivers (to do two things: quiesce the device before the change, and adjust the frequency after the change). Deciding which frequency to use is the responsability of code from somewhere else (either cpufreq governors or userspace).

The most recent version of the code can be found at http://repo.or.cz/w/linux-2.6/s3c2410-cpufreq.git on the branches s3c2410-cpufreq (for code which applies on top of the upstream "git HEAD" kernel) and s3c2410-cpufreq-gta01 (for code which applies on top of the OpenMoko patched kernel).

Contents

To build

NOTE: Do not run on real hardware, unless you really know what you are doing. This code is completely untested. It might even not compile.


You need a fully built OpenMoko tree for the cross-compiler and uboot tools.

  1. Checkout the s3c2410-cpufreq-gta01 branch
  2. Copy the defconfig file from branches/src/target/kernel/2.6.24.x/config on the OpenMoko svn as the .config file
  3. Set the PATH to the correct value (check the run.* temporary files bitbake generates to find out the correct value)
  4. Run make ARCH=arm CROSS_COMPILE=arm-angstrom-linux-gnueabi- oldconfig and answer correctly the questions
  5. Run make ARCH=arm CROSS_COMPILE=arm-angstrom-linux-gnueabi- uImage
  6. The last line of the output tells you the correct file which should be flashed on the emulator.

Hints

  • Enable CONFIG_CPU_FREQ_DEBUG and add cpufreq.debug=7 to the kernel command line to enable the relevant debug output. You can also enable it after boot by going to the correct place at /sys/modules and changing the values there.
  • To test on 2.6.22.5, use git format-patch -o dir master..s3c2410-cpufreq to generate a set of patches, and apply them all on top of the OpenMoko patches 2.6.22.5 kernel, except the framebuffer patch(es) (the 2.6.24.x code for the framebuffer is different from the 2.6.22.5 one).

What's already done

S3C2410A cpufreq driver

This code resides on arch/arm/mach-s3c2410/s3c2410-cpufreq.c and is responsible for actually changing the frequency.

The list of available frequencies and their parameters (HCLK/PCLK dividers) can be found on this file.

Timer 4 driver

Timer 4 is used by the kernel as the periodic tick timer. The cpufreq notifier on arch/arm/plat-s3c24xx/time.c is responsible for adjusting the current and reload values of the timer to match the new frequency.

Serial driver

The cpufreq notifier on drivers/serial/s3c2410.c is responsible for (if possible) pausing both sides of the serial transmission before the frequency change and reloading the baud generator (and unpausing the serial transmission) after the change.

Pausing the transmission is currently only possible with hardware flow control. Since both serial ports on GTA01 use hardware flow control, that's not a problem (unless you are using the serial console).

Framebuffer driver

The cpufreq notifier on drivers/video/s3c2410fb.c is responsible for turning off the video before the frequency change and reloading its configuration after the frequency change.

NAND driver

The cpufreq notifier on drivers/mtd/nand/s3c2410.c is responsible for reprogramming the NAND timings after a frequency change.

What's missing

S3C2410A cpufreq driver

The frequency table should be checked for better divisor values (1:1:1 might be appropriate in more situations), and the code should be able to deal with HDIVN1 and different bus modes.

Timer 4 driver

The cpufreq notifier should perhaps use a higher priority, so as to run before all the others for more precise timekeeping.

NAND driver

The cpufreq notifier should make sure the NAND is quiesced before the frequency change, or change the timings to the most conservative ones before the frequency change.

Missing drivers

  • Backlight (uses a PWM timer; not a problem with full brightness)
  • Vibrator (uses a PWM timer)
  • MMC/SD (can be avoided by not plugging it in)
  • I2C (without a notifier and booting at the maximum speed, it can only get slower; there should be no side effects, since there's no minimum bus speed)
  • IIS
  • SPI

Frequency restrictions

Some drivers need specific frequencies. Currently there's no way to tell the cpufreq core of that, but at minimum a cpufreq notifier should be used to turn them off if they won't be able to work with that frequency.

  • The datasheet says the USB device gets unstable if the frequency is below a minimum value
  • When using the IIS device, changing the frequency might not be a good idea, since it can cause audio glitches. It also probably needs a minimum frequency depending on the output audio frequency and sample size.
  • The LCD device might also need a minimum frequency.
  • The serial baudrate generator might not have a good enough divider value for some baud rates and some clock frequencies (for instance, 115200 at 12000 kHz)

Testing report

2008-01-19

2.6.24.x doesn't boot on GTA01. I hand-extracted the patches and applied them all on top of GTA01's 2.6.22.5 kernel, except for the framebuffer patch (which depends on some changes made after 2.6.24). It was booted on a GTA01 without the SD card or the SIM attached (just in case). The results were:

  • Setting to 202800 kHz: lost the USB connection. It would be recovered only after rebooting (setting the frequency back to 266000 wasn't enough). The rest of the tests were made with the on-screen terminal.
  • Setting to 133000 kHz (actually 135000 kHz): noticeable flicker, possibly because the framebuffer patch wasn't applied. The timer seems to be running at exactly half the correct rate (running "date" twice with 10s delay showed the GTA01's clock advanced only 5s).
  • Rebooting (after coming back to 266000 kHz) fixed the USB connection.

2008-01-20

Same 2.6.22.5 kernel, with several bugfixes applied (in particular, the problem which prevented the timer notifier from running was fixed).

This time, the whole test was done from the terminal, to avoid the USB lockup.

  • Setting to 133000 kHz (actually 135000 kHz): same flicker. Plugging briefly the USB shows it to still be alive. Time runs at the correct speed.
  • Setting to 100000 kHz (actually 101250 kHz): even worse flicker.
  • Setting to 12000 kHz (SLOW mode on): the flicker gets psychedelic (colors start oscillating). VERY slow (taking about a second to register a "key" press).
  • Back to 266000 kHz: works normally. Plugging briefly the USB again shows it to still be alive. Time still running at correct speed.

2008-01-22

After a few more bugfixes, still 2.6.22.5, also tested from the terminal (untethered).

This time, due to the nature of the tests, I commented out all speeds below 12000 kHz.

  • Setting governor to "conservative": works, and the flicker patterns shows neatly the change of operating frequencies.
  • Setting governor to "ondemand": works, however after a few seconds on 12000 kHz it got "stuck" there. For some reason, cat scaling_cur_speed locked up (not answering to Ctrl-C/Ctrl-Z); requested a shutdown via the power menu, which worked (still with the severe flicker showing it was still at 12000 kHz.

Another test was to see if gsmd still worked.

  • With "conservative": timeout.
  • With "userspace", testing by changing the frequency and typing "at" at libgsmd-tool -m atcmd and checking for the response: worked at 266000, 133000, 66000, 33000, and failed at 12000 (timeout).

2008-01-26

Trying to find out why gsmd stops working at 12000.

  • Adding a debug printk to the end of s3c24xx_serial_getclk, the error is 7% with 12000, while below 2% at the other frequencies tested above.
  • Trying to make gsmd run at 57600 instead of 115200 didn't work, even after making the kernel use the same speed for the serial console. Changing from 57600 to 115200 worked without even rebooting; looks like the modem doesn't autodetect the serial speed.

See also

Personal tools

The cpufreq subsystem is a Linux kernel subsystem responsible for managing the CPU frequency. It's commonly used in laptops to reduce power usage when idle.

On the S3C2410, adjusting the CPU frequency changes the clocks for almost all devices on the chip. This means that, to write a cpufreq driver for the S3C2410, you have to also adjust some values on almost all devices.

This project aims to implement the cpufreq driver (which manages the CPU frequency transitions) and a cpufreq notifier for all the affected drivers (to do two things: quiesce the device before the change, and adjust the frequency after the change). Deciding which frequency to use is the responsability of code from somewhere else (either cpufreq governors or userspace).

The most recent version of the code can be found at http://repo.or.cz/w/linux-2.6/s3c2410-cpufreq.git on the branches s3c2410-cpufreq (for code which applies on top of the upstream "git HEAD" kernel) and s3c2410-cpufreq-gta01 (for code which applies on top of the OpenMoko patched kernel).

To build

NOTE: Do not run on real hardware, unless you really know what you are doing. This code is completely untested. It might even not compile.


You need a fully built OpenMoko tree for the cross-compiler and uboot tools.

  1. Checkout the s3c2410-cpufreq-gta01 branch
  2. Copy the defconfig file from branches/src/target/kernel/2.6.24.x/config on the OpenMoko svn as the .config file
  3. Set the PATH to the correct value (check the run.* temporary files bitbake generates to find out the correct value)
  4. Run make ARCH=arm CROSS_COMPILE=arm-angstrom-linux-gnueabi- oldconfig and answer correctly the questions
  5. Run make ARCH=arm CROSS_COMPILE=arm-angstrom-linux-gnueabi- uImage
  6. The last line of the output tells you the correct file which should be flashed on the emulator.

Hints

  • Enable CONFIG_CPU_FREQ_DEBUG and add cpufreq.debug=7 to the kernel command line to enable the relevant debug output. You can also enable it after boot by going to the correct place at /sys/modules and changing the values there.
  • To test on 2.6.22.5, use git format-patch -o dir master..s3c2410-cpufreq to generate a set of patches, and apply them all on top of the OpenMoko patches 2.6.22.5 kernel, except the framebuffer patch(es) (the 2.6.24.x code for the framebuffer is different from the 2.6.22.5 one).

What's already done

S3C2410A cpufreq driver

This code resides on arch/arm/mach-s3c2410/s3c2410-cpufreq.c and is responsible for actually changing the frequency.

The list of available frequencies and their parameters (HCLK/PCLK dividers) can be found on this file.

Timer 4 driver

Timer 4 is used by the kernel as the periodic tick timer. The cpufreq notifier on arch/arm/plat-s3c24xx/time.c is responsible for adjusting the current and reload values of the timer to match the new frequency.

Serial driver

The cpufreq notifier on drivers/serial/s3c2410.c is responsible for (if possible) pausing both sides of the serial transmission before the frequency change and reloading the baud generator (and unpausing the serial transmission) after the change.

Pausing the transmission is currently only possible with hardware flow control. Since both serial ports on GTA01 use hardware flow control, that's not a problem (unless you are using the serial console).

Framebuffer driver

The cpufreq notifier on drivers/video/s3c2410fb.c is responsible for turning off the video before the frequency change and reloading its configuration after the frequency change.

NAND driver

The cpufreq notifier on drivers/mtd/nand/s3c2410.c is responsible for reprogramming the NAND timings after a frequency change.

What's missing

S3C2410A cpufreq driver

The frequency table should be checked for better divisor values (1:1:1 might be appropriate in more situations), and the code should be able to deal with HDIVN1 and different bus modes.

Timer 4 driver

The cpufreq notifier should perhaps use a higher priority, so as to run before all the others for more precise timekeeping.

NAND driver

The cpufreq notifier should make sure the NAND is quiesced before the frequency change, or change the timings to the most conservative ones before the frequency change.

Missing drivers

  • Backlight (uses a PWM timer; not a problem with full brightness)
  • Vibrator (uses a PWM timer)
  • MMC/SD (can be avoided by not plugging it in)
  • I2C (without a notifier and booting at the maximum speed, it can only get slower; there should be no side effects, since there's no minimum bus speed)
  • IIS
  • SPI

Frequency restrictions

Some drivers need specific frequencies. Currently there's no way to tell the cpufreq core of that, but at minimum a cpufreq notifier should be used to turn them off if they won't be able to work with that frequency.

  • The datasheet says the USB device gets unstable if the frequency is below a minimum value
  • When using the IIS device, changing the frequency might not be a good idea, since it can cause audio glitches. It also probably needs a minimum frequency depending on the output audio frequency and sample size.
  • The LCD device might also need a minimum frequency.
  • The serial baudrate generator might not have a good enough divider value for some baud rates and some clock frequencies (for instance, 115200 at 12000 kHz)

Testing report

2008-01-19

2.6.24.x doesn't boot on GTA01. I hand-extracted the patches and applied them all on top of GTA01's 2.6.22.5 kernel, except for the framebuffer patch (which depends on some changes made after 2.6.24). It was booted on a GTA01 without the SD card or the SIM attached (just in case). The results were:

  • Setting to 202800 kHz: lost the USB connection. It would be recovered only after rebooting (setting the frequency back to 266000 wasn't enough). The rest of the tests were made with the on-screen terminal.
  • Setting to 133000 kHz (actually 135000 kHz): noticeable flicker, possibly because the framebuffer patch wasn't applied. The timer seems to be running at exactly half the correct rate (running "date" twice with 10s delay showed the GTA01's clock advanced only 5s).
  • Rebooting (after coming back to 266000 kHz) fixed the USB connection.

2008-01-20

Same 2.6.22.5 kernel, with several bugfixes applied (in particular, the problem which prevented the timer notifier from running was fixed).

This time, the whole test was done from the terminal, to avoid the USB lockup.

  • Setting to 133000 kHz (actually 135000 kHz): same flicker. Plugging briefly the USB shows it to still be alive. Time runs at the correct speed.
  • Setting to 100000 kHz (actually 101250 kHz): even worse flicker.
  • Setting to 12000 kHz (SLOW mode on): the flicker gets psychedelic (colors start oscillating). VERY slow (taking about a second to register a "key" press).
  • Back to 266000 kHz: works normally. Plugging briefly the USB again shows it to still be alive. Time still running at correct speed.

2008-01-22

After a few more bugfixes, still 2.6.22.5, also tested from the terminal (untethered).

This time, due to the nature of the tests, I commented out all speeds below 12000 kHz.

  • Setting governor to "conservative": works, and the flicker patterns shows neatly the change of operating frequencies.
  • Setting governor to "ondemand": works, however after a few seconds on 12000 kHz it got "stuck" there. For some reason, cat scaling_cur_speed locked up (not answering to Ctrl-C/Ctrl-Z); requested a shutdown via the power menu, which worked (still with the severe flicker showing it was still at 12000 kHz.

Another test was to see if gsmd still worked.

  • With "conservative": timeout.
  • With "userspace", testing by changing the frequency and typing "at" at libgsmd-tool -m atcmd and checking for the response: worked at 266000, 133000, 66000, 33000, and failed at 12000 (timeout).

2008-01-26

Trying to find out why gsmd stops working at 12000.

  • Adding a debug printk to the end of s3c24xx_serial_getclk, the error is 7% with 12000, while below 2% at the other frequencies tested above.
  • Trying to make gsmd run at 57600 instead of 115200 didn't work, even after making the kernel use the same speed for the serial console. Changing from 57600 to 115200 worked without even rebooting; looks like the modem doesn't autodetect the serial speed.

See also