weird name for a kernel..or for pretty much anything..dude
From the man himself:
[Kernel][CWM]Fecality I9100 3/7/2012 (GPU Voltage Fix)
- Thread starter Entropy512
- Start date
From the man himself:
Pretty consistent with my results. Once you get rid of the XXKI3 wifi PM bug and the BT-AMP bug in XWKK5, you can get amazing life when on wifi. Cell data is a different story - I see drain as high as around 1.5%/hour when my phone is at my desk (very weak signal, so the radio cranks up the power. Nothing the kernel can do about that.
)Another interesting thing - when on wifi and a permanent wakelock is held (try the Wake Lock app), drain is still only around 1.5%/hour.
Just flashed with no problems and under volted it a bit. . When flashing i saw it say something about cognition. Is he still about? Design gears is an epic dev! Cognition was the best rom about! No questions asked!
Sent from my GT-I9100 using xda premium
. When flashing i saw it say something about cognition. Is he still about? Design gears is an epic dev! Cognition was the best rom about! No questions asked! Sent from my GT-I9100 using xda premium
Pretty consistent with my results. Once you get rid of the XXKI3 wifi PM bug and the BT-AMP bug in XWKK5, you can get amazing life when on wifi. Cell data is a different story - I see drain as high as around 1.5%/hour when my phone is at my desk (very weak signal, so the radio cranks up the power. Nothing the kernel can do about that.
Another interesting thing - when on wifi and a permanent wakelock is held (try the Wake Lock app), drain is still only around 1.5%/hour.
I'm on XWLA4. About performance in benchmarks, CF-Bench reported above-average results. Battery life testing will continue, tomorrow I'll report on battery life with detailed BBS logs and what-not. Anything you're trying to pin down we can help find?
Edit: there's a strange bug when unlocking my phone? Sometimes the unlock ring won't appear when I touch the screen. It won't work until I repeatedly lock and unlock the screen. I suspect it has to do with the kernel, but I have been tinkering with vitalij's touchscreen fix. Don't know if it relates, but still.
Sent from my GT-I9100 using XDA App
Last edited:
He was on the I777 for a while - My kernel flash zips thank him for the information on flashing a kernel on an I777. That's been there since October, when the I777 launched. (Flashing a kernel on the Infuse was, honestly, a lot uglier, hence the "no more redbend_ua makes Entropy happy".)Just flashed with no problems and under volted it a bit.
Sent from my GT-I9100 using xda premium
He left the I777 after 2-3 months for the GNex - he wanted to graduate from smali hacks/mods to full-blown source development. (I don't blame him!)
Hmm... That's REALLY strange. A bunch of people reported that when I first went from the I777 codebase to I9100 update3 - All of those reports disappeared when I reverted the touchscreen driver to the I777 one. GM did the same revert with Siyah.
Hmm... However, I may have maintained the I777 #iddefs, while GM might not have. I think that when building for I9100, it may use the I9100 code still. I'll look into this tonight.
Right now, I do not know of any battery life issues in this kernel that affect XWKL1 or XWLA4. The only power management regression I ever had when going from I777 to I9100 update3 was specific to XXKI3, and I now have code that prints alerts to dmesg when that problem occurs.
I've primarily been focusing on standby/idle/lowload power management lately, but I may start looking at higher-load PM.
The problem is that there is often not that much can be done - I question whether the arm topology/sched_mc patches do much on a dual-core unit - especially with hotplugging involved. They'll matter much more on a quadcore CPU. I haven't written them off yet, but they've been lower priority due to their questionable benefit (especially when I read them - they seem to contain lots of infrastructure/groundwork for the future, but without actually making significant changes.)
Right now I'm in the process of slowly bringing things back to the I777 2/7 baseline - That release had stability issues for some people so I reverted most patches out and have been readding one at a time, even though I think I already identified the problem patch. Once that's done - I'll look into maybe changing minimum display brightness and CWM Manager 3.0 compatibility.
The problem is that there is often not that much can be done - I question whether the arm topology/sched_mc patches do much on a dual-core unit - especially with hotplugging involved. They'll matter much more on a quadcore CPU. I haven't written them off yet, but they've been lower priority due to their questionable benefit (especially when I read them - they seem to contain lots of infrastructure/groundwork for the future, but without actually making significant changes.)
Right now I'm in the process of slowly bringing things back to the I777 2/7 baseline - That release had stability issues for some people so I reverted most patches out and have been readding one at a time, even though I think I already identified the problem patch. Once that's done - I'll look into maybe changing minimum display brightness and CWM Manager 3.0 compatibility.
On hotplugging: http://xdaforums.com/showpost.php?p=22460365&postcount=244
I'm not sure if it's really relevant because it's from a GN developer, but hey.
I can confirm that I've hit this bug a couple times, and hadn't with Siyah in at least a month. Would getting you a dmesg help?
I've just turn-on & off my screen for about 30 times and i always have lockscreen's ring, on stock LA4 atm.
Clear RAM in the task manager (wakelocks), let the phone sleep for 15 minutes and try again. It's usually when it happens for me, I think Deep Sleep is the problem. When the phone is plugged in though, everything works just fine.
Last edited:
The OMAP4 has very different architecture than ours.
For one, his comment that he saw 10 hotplug events in one second is simply not possible on our device - our hotplugging timer runs at 2 Hz. This allows for "ramp before hotplug" in most situations.
For another - on our device, having CPU1 hotplugged in blocks LPA and AFTR idle modes.
Give me some time for a code review - This one might be simple (reworking the #ifdefs), and no one was able to catch anything obvious with a dmesg on the I777.
And yes, I'm 90% positive it was only on wakeup from deep sleep.
The problem was intermittent - some people (including myself) never experienced it, others had it occasionally, one guy had it happen all the time.
It also affected the ability to touch the screen during phone calls for him.
It also affected the ability to touch the screen during phone calls for him.
Similar threads
- Locked
Top Liked Posts
-
There are no posts matching your filters.
-
37Read the Known Issues section of this post below along with the FAQ a few posts down before posting.
A few people have asked about I9100 builds of my Daily Driver kernel for the SGH-I777. I normally hate blind development, but since most I9100 guys are doing it for I777 targets, I may as well try going the other way. Since it fails the definition of being a "daily driver" as I own an I777 and not I9100 - it will not be using the Daily Driver name.
This should be compatible with most stock-derived Gingerbread firmwares, with a few exceptions listed below. It is NOT compatible with CM7/MIUI or any other AOSP-derived firmware. It is NOT compatible with ICS and WILL NOT BE until ICS kernel source for the I9100 is released. At that point a new thread will be created for those kernels. I am testing it currently with self-deodexed/debloated/Hellraised XWLA4.
The following firmware bases are likely to deliver poor battery life with this kernel due to wifi bugs in the firmware:
XXKI3, UCKK6, XWKK5 - especially UCKK6 and XWKK5
This kernel series is intended to be similar in spirit to my Daily Driver series for the Infuse at http://xdaforums.com/showthread.php?t=1212795
It is built from sources at https://github.com/Entropy512/linux_kernel_sgh-i777/commits/oc_codeworkx, and initramfs at https://github.com/Entropy512/initramfs_sgh-i777/commits/master
My general goals are to focus on stability and battery life. If it comes to a tradeoff between performance and the above two, I will choose stability/battery life. In general I will choose stability first, with the exception of undervolting.
I keep my initramfs clean - There are no tweaks or other automated scripts built into this kernel. For a collection of tweaks I am working on and documentation on them, see http://xdaforums.com/showthread.php?t=1378080
Current features:
- I9100 Update3 base
- Touchscreen drivers from I777 source code base
- codeworkx's cpuidle patch - should improve battery life a bit. In most cases it will likely not improve things much, but in rare cases it will result in significant improvements. (I only have one partially-reproducible test case on the Infuse so far)
- JHash 3
- BFQ I/O scheduler
- CIFS module in initramfs
- CWM 5.0.2.8 compiled from CM7 sources on 2/28/2012
- "insecure" kernel (meaning root in ADB)
- CPU governor set to Conservative by default to conserve some battery - this will make your device slightly less responsive, use SetCPU or a similar app to return to ondemand if you want it, or reduce the conservative polling interval
- Filesystem readahead tweaks in initramfs
- netarchy's Sleep of Death fix
- netarchy's conservative governor tuning patch - should improve responsiveness of devices when using the conservative governor if you reduce the polling interval (misnamed as sample_rate) - the I9100 community calls this "lionheart" even though it's really only a 2-line patch
- Battery charge current monitoring (CurrentWidget) support - only reports charge current and not discharge, and reports a value 2.85 times the actual current. Use CurrentWidget's "operation on value" to divide by 2.85.
- Miscellaneous bugfixes pulled from Ninphetamine and CM7 sources - see github for details
- /system/etc/init.d support in initramfs - Note that this only runs stuff in /system/etc/init.d - ROM developers or you need to create it. Attached is an example script that will change the CPU frequency governor to ondemand if placed in /system/etc/init.d and set to executable
- Four "use at your own risk" features that trade performance for stability - See Post #4 for details
- Standard bootanimation support
- /proc/last_kmsg crash debugging support
- NFS modules in initramfs - note that they must be insmodded in a specific order: sunrpc.ko, lockd.ko, then nfs.ko
- Fix for fuel_alerted perma-wakelocks
- Fix for wifi tethering on I9100 ROMs that have been Hellraised to I777
- Bump up TCP buffer sizes in initramfs to match that of the Infuse - may help network performance in some cases
- cpuidle driver from Tab 7 Plus kernel - allows entry into AFTR more often - includes MFC interoperability fix (Should not break video playback.)
- Support Bluetooth HID on newer firmware bases
- 3-step GPU clock/voltage control
- Extended hotplug tuning
- Support for Xan's ExTweaks universal tuning app - https://market.android.com/details?id=com.darekxan.extweaks.app
Planned features, short term:
- Determine if issues with PINlock when running I9100 firmwares on I777 have kernel involvement, and fix them if this is the case.
Planned features, mid-term:
- Potentially my first departure from "standard" governors, with a focus on improving hotplugging logic
Planned features, long-term:
- Improved battery charge algorithm for faster charging - Initial research indicates we have an alternate battery charger chip (MAX8922) that differs from the MAX8997 used in the I9100. We DO have an 8997 also - but on our device for some reason Samsung decided to use an alternate chip instead of using the 8997's built-in charging. This means we have far fewer options (90,400,660 mA) in terms of charge rates compared to the I9100 (from 200 to 950 in 50 mA steps). So we might not be able to implement any fancy charging algorithms.
Features not planned:
- BFS process scheduler - I have only once ever seen a test case where this clearly outperformed the mainline Linux scheduler (multithread x264 encoding) - The mainline schedule was fixed in the next release and BFS now has no performance benefits
- Any feature that trades off stability or data integrity for performance unless it can be disabled entirely and defaulted to "off"
- Any feature that cannot have functionality tested without a paid app. Interface-only checks don't cut it - I don't want users complaining that the app they paid for didn't work because an interface check worked but function didn't
- Touch recovery - too prone to unintentional user input in dangerous situations. I will reconsider this with ICS.
- ARM_TOPOLOGY/SCHED_MC - See http://xdaforums.com/showpost.php?p=23193259&postcount=100 for details
Known issues:
- Power management regression somewhere between 12/8/11 and 1/2/12 - Intermittent high drain without high AOS or reduced deep sleep percentage when on some wifi networks - seems more likely if GPS is used when connected to wifi. Wifi with high AOS/reduced deep sleep is not a kernel problem. This appears to only happen on some firmwares - it happens on XXKI3 but not XWKL1. It is likely connected to a wifi power management bug in some firmwares. A debugging feature in 2/7 and later will allow identification of such firmwares - see http://xdaforums.com/showpost.php?p=22581928&postcount=1777 for details
- Some people have reported touchkey lights becoming disabled until the screen is turned off and back on again. Under investigation - seems to mainly happen on firmwares with BLN-modded liblights even if the BLN app isn't used
- Internal and External SD card are swapped in CWM currently
- Keymap is a little funky in CWM for I9100 users. VolUp/VolDn to move, Power to select is reported to work, but bottom keys are funky. Working on it
- A few users have reported that when Hellraising I9100 firmwares to I777, they are unable to set up a PIN-lock. If this occurs for I9100 users on their own devices, I need to look into initramfs fixes and this will be my top priority.
(Tested on Linux, not tested MacOS/Windows but should work) Heimdall - Extract the contents of the tar file, enter download mode, and flash with the following command line:
Code:heimdall flash --kernel zImage
Flashing instructions for .zip releases:
Flash in CWM, or extract the zImage and use the Heimdall instructions above.
Please do not ask how to enter download mode or install Heimdall/Odin in this thread - these are basic generic skills anyone flashing custom firmwares on Samsung devices should know and plenty of documentation exists elsewhere. If you really need to ask, use the General forum, or if created, the Q&A forum. I want to try to keep this thread clean and only with bug reports and issues specific to this release, not general HOWTO or troubleshooting posts. Some of the information you need is in jivy26's FAQs thread at http://xdaforums.com/showthread.php?t=1288112 - Reading at least the first post of this thread in its entirety is STRONGLY recommended.
Bug reports:
If you have a crash (reboot all the way to Galaxy S I9100 screen), use ADB dump the contents of /proc/last_kmsg and post
If you have oddball behavior, include a clearly reproducible test case with your report, or use ADB to obtain a dmesg and logcat capturing the odd behavior at the time of error.
Similar to flashing - using ADB and obtaining last_kmsg, dmesg, and logcat dumps are basic skills that anyone working with custom firmwares on Android devices should have. If you need help with these, do some searching, or post in the General forum or Q&A forum.
Firmware ("ROM") Developers:
While I cannot restrict anyone from putting this kernel into a ROM as long as links are given to the github sources for GPL compliance, I request that anyone who includes this kernel in a firmware release does the following out of courtesy:
Link to this thread
Clearly indicate in your firmware changelog which Daily Driver kernel release is included in your firmware release whenever you change DD releases - this lets users identify whether a fix is present in the kernel they're using or not
Kernel Developers:
Similar to my request for ROM developers, while I can't restrict you from doing anything, I ask as a courtesy that if you cherry-pick my commits, you do the following:
Please don't rebase my commits into a large multi-feature without consulting me - rebasing related bugfixes together is OK.
Please try not to implement lots of unrelated features or bugfixes in a single git commit - it makes it hard to reimplement that when Samsung drops new sources or releases a new device
ALL OF MY RELEASES ARE NAMED BY RELEASE DATE - MMDDYYYY. See the changelog for differences between Experimental (exp) and non-exp versions for days where dual releases are made.9I'm not going to mess with sched_mc right now as I don't think it'll do anything - that post was my reasoning why.
There's just not much you can do power-management wise as far as multicore scheduling tricks when you only have clock gating available to you (which is the only thing available when the second core is on). Two cores at 50% will have the same consumption as one at 100 and one at 0, for example. If the cores could hit AFTR or LPA this would be different - the 100/0 combo would be better. (Kind of like hotplugging but without the power penalty involved with the actual hotplug event.)
So screen-on power consumption is likely best handled by advanced GPU clocking (You're way ahead of me there...) and changing the hotplugging logic (Probably some neat tricks to be done here, even after ICS drops).
Could be different in ICS though. Sammy might add some undocumented idle features. (Idle states of Exynos are COMPLETELY undocumented in the TRM... Other than "we have them"...) See the Tab 7 Plus cpuidle backport - it can hit AFTR far more often now. If there's a way to hit AFTR or LPA when the second core is on - everything changes.
What I'm thinking of experimenting with is tying the hotplug logic into a governor - bias the hotplug decisions based on frequency and ensure decisions are synced with load. Favor singlecore heavily up to 800 MHz or so, then bias more towards dualcore past this, since the voltage starts climbing more per frequency step past 800.
Something like this, a version of conservative where, instead of the following logic at 800 MHz:
If (load > up_threshold) go to 1000
it would be:
If (load > up_threshold AND only one core on AND we have multiple threads) hotplug the second core instead of going to 1000.
Not sure whether there are any good ways to determine if our load is highly multithreaded or single threaded. (This is another thing that may work better on ICS, since it supposedly is more multithreaded.)
I did not fully read the post and I thought you would implement it. sorry I misunderstood.
then I will read it now
by the way... why is this thread rated 4?
as an analogy: there are several contests on tv in my country, mainly singing and talent contests...
elections of the talent contests takes about 4 times than singing contests which makes me think that there are 4 times more clowns in my country than talents.
this thread is maybe the most original and development related thread lately and it doesn't get 5 from users which makes me think that there are more "users" (I'm being polite) than developers in this forum.
anyway.. let's think positive and hope that 4 points will make the noobs stay away.8Change Log
The primary change log will be kept in my I777 kernel thread at http://xdaforums.com/showpost.php?p=18129037&postcount=2
Unless I9100-specific items are implemented, this post will simply indicate which I777 release the kernel corresponds to.
3/7/2012:
Corresponds to Daily Driver 3/7/2012
03/05/2012:
Corresponds to Daily Driver 3/5/2012
03/04/2012-B:
Corresponds to Daily Driver 3/4/2012-B
03/04/2012:
Corresponds to Daily Driver 3/4/2012
03/03/2012-B:
Corresponds to Daily Driver 3/3/2012-B
03/03/2012:
Corresponds to Daily Driver 3/3/2012
Initial release was bad, that was an I777 build. Reuploaded at 4:22 PM Eastern Standard time on 3/3 - Redownload if you downloaded before that
3/02/2012:
Corresponds to Daily Driver 3/02/2012
2/28/2012:
Corresponds to Daily Driver 2/28/2012
IMPORTANT: Internal/External sdcards have been swapped in CWM to be consistent with Android standards!
2/26/2012:
Corresponds to Daily Driver 2/26/2012
Initial release, 02/15/2012:
I know it's dated 2/15 and not 2/26, because that's when it was built. Thanks to khartaras for indicating that it isn't at least a horrible turd when used on an I9100.
This release corresponds to the 2/15/2012-C release of I777 Daily Driver7Here Be Dragons
This post is for features present in the kernel that are "use at your own risk" - They have either potential or guaranteed negative side effects if used.
Overclocking (CPU):
Enable using SetCPU or a similar app
USE AT YOUR OWN RISK. DO NOT REPORT BUGS OR PROBLEMS IF YOU ARE OVERCLOCKING OR UNDERVOLTING. IF YOU EXPERIENCE ANY STABILITY PROBLEMS, DISABLE ALL OC/UV
Overclocking (GPU):
See Ninphetamine kernel for documentation - Same control method
USE AT YOUR OWN RISK. DO NOT REPORT BUGS OR PROBLEMS IF YOU ARE OVERCLOCKING. IF YOU EXPERIENCE ANY STABILITY PROBLEMS, DISABLE ALL OC
Per-File fsync() disable:
This allows you to disable per-file write forced syncs. (e.g. if an app tries to force a write straight to disk, it'll just go to cache). This achieves the same goal as the modded sqlite hacks seen in tweaks such as USAS, however it can be disabled at runtime.
WARNING: THIS CAN CAUSE DATA LOSS OR CORRUPTION IN A CRASH
To enable, do the following in a terminal, or add it to an init.d script (look at my ondemand script as an example):
Code:echo "1" > /sys/module/sync/parameters/fsync_disabled
And to disable (return to the default):
Good for around 200 points of epeen in the database benchmarks in Antutu or 500-600 points of epeen in Quadrant. Real-world benefit: Probably not worth the data integrity risk, but you've got a choice now.Code:echo "0" > /sys/module/sync/parameters/fsync_disabled
Backlight Notifications (BLN):
This allows the touchkey backlights to be used for notifications. Some stock apps (such as stock MMS) don't support it. Supposedly services.jar mods can change this.
This WILL drain your battery when a notification is active due to a wakelock that holds deep sleep. Sorry, it's either this or instability for the time being.
In addition to the BLN control app, the ROM needs a modified liblights file for this to work
Attached here - Liblights - both BLN-modified (extracted from VillainROM 3.0) and stock I777
To install, take the file and push it to /system:
Then rebootCode:adb remount adb push <file> /system/lib/hw/lights.SGH-I777.so adb chmod 644 /system/lib/hw/lights.SGH-I777.so
Note that on a Hellraised ROM, you need to replace SGH-I777 with GT-I9100. This includes manually ported ROMs like Cognition 777
Like my prerooted system image, this file is compressed using 7-Zip to prevent people from trying to flash it with CWM6A bit on SCHED_MC/ARM_TOPOLOGY
One patchset that has been floating around for a while is the ARM_TOPOLOGY/SCHED_MC patchset that is supposed to improve battery life.
My first few readthroughs of the code didn't lead me to believe that there was that much improvement to be had, so it has been a low priority for a while. I took a further look, and have decided not to implement this patchset.
First, the background on the topology/sched_mc patches is at https://wiki.linaro.org/WorkingGroups/PowerManagement/Specs/sched_mc
One important thing to note - Most of the stated goals of ARM_TOPOLOGY are to improve the kernel's awareness of how CPUs are grouped on hyperthreaded/multicore/multipackage systems. The thing is, on systems with simple topologies such as a dualcore system without hyperthreading, ARM_TOPOLOGY is redundant - the kernel already knows there are two cores on a single package that share the same clock.
So that leaves SCHED_MC - This offers multiple choices for scheduling tasks between two CPUs. The first is a basic load balancer that tries to keep load even between CPUs, and is multipackage/hyperthreading aware. We don't have hyperthreading or multipackage, so this mode offers little to no advantage over the default scheduling methods.
The second is one that, instead of trying to keep load even, tries to keep load on one core until it is fully loaded, so that the second core can enter deeper IDLE states. The assumption here is - it is better to allow the second core to enter very deep idle states and the other core doesn't idle much at all rather than have both cores at lower loads and not be able to hit the deeper idle states.
This is where the problem with SCHED_MC on our device lies - When the second core is online, all of the deeper idle states (AFTR, LPA) are locked out. This leaves only basic clock gating (IDLE), which is fairly fine grained. So there is no advantage to asymmetricaly loading the cores so that one can hit deeper idle states, because there are no deeper states to hit when the second core is hotplugged in. See https://github.com/Entropy512/linux_kernel_sgh-i777/blob/master/arch/arm/mach-s5pv310/cpuidle.c#L990 (and also line 1016)
Based on what I've seen of OMAP4 source code - this is the same on OMAP4 devices like GNex. On OMAP4, deeper idle states can be hit when both cores are online BUT both cores must agree on the idle state - Thus it is better to have the cores similarly loaded (so they can both enter idle) than to have asymmetric load (because one core entering idle will block deeper idle states on the other core, if I am reading the code correctly.) I could be wrong about this on OMAP4 - however I am absolutely positive that deep idle states are blocked on Exynos when the second core is online.
So this leaves us only with improved hotplugging logic to try and save power - which I'm going to look into once I finish a few other items after PatchADay is complete.
(And yeah, I know it has proven to not quite be one patch per day - but it has been one patch per release except for reverts and the CWM stuff.)
