[PATCH 2/2] arc770: move arc patches to taregt/linux/generic

[PATCH 2/2] arc770: move arc patches to taregt/linux/generic

[Alexey Brodkin]

Hi Felix,

On Fri, 2016-01-15@11:45 +0100, Felix Fietkau wrote:
> On 2016-01-15 00:12, Alexey Brodkin wrote:
> > Given those patches are relevant to any ARC platform and even
> > ISA version it makes perfect sense for patches to exist
> > in one place instead of being duplicated for each new ARC-based ASIC.
> > 
> > Note this is a prerequisite for upstreaming of ARC HS38 support in
> > OpenWRT.
> > 
> > Signed-off-by: Alexey Brodkin <abrodkin at synopsys.com>
> > Cc: Felix Fietkau <nbd at openwrt.org>
> > Cc: Jo-Philipp Wich <jow at openwrt.org>
> > Cc: Jonas Gorski <jogo at openwrt.org>
> > --- /dev/null
> > +++ b/target/linux/generic/patches-4.4/331-arc-remove-dependency-on-DEVTMPFS.patch
> > @@ -0,0 +1,36 @@
> > +From adfbf9e6cad93281cffceab078e7f6f2a8e094f9 Mon Sep 17 00:00:00 2001
> > +From: Alexey Brodkin <abrodkin at synopsys.com>
> > +Date: Thu, 13 Aug 2015 01:56:02 +0300
> > +Subject: [PATCH 1/2] openwrt: arc - remove dependency on DEVTMPFS
> > +
> > +OpenWRT builds initramfs so that it doesn't require DEVTMPFS so dropping
> > +this dependency. That helps to escape 2 separate kernel rebuilds with
> > +and without initramfs.
> > +
> > +2 builds happen because OpenWRT first builds kernel and later modules.
> > +When building entire kernel with simple "make" INITRAMFS sets to a real
> > +value and so was triggering DEVTMPFS selection. Then when building only
> > +modules with "make modules" command INITRAMFS is zeroed and so kernel
> > +config was changing that lead to full kernel rebuild.
> > +
> > +Signed-off-by: Alexey Brodkin <abrodkin at synopsys.com>
> > +---
> > + arch/arc/Kconfig | 2 --
> > + 1 file changed, 2 deletions(-)
> > +
> > +diff --git a/arch/arc/Kconfig b/arch/arc/Kconfig
> > +index 6312f60..a95bab3 100644
> > +--- a/arch/arc/Kconfig
> > ++++ b/arch/arc/Kconfig
> > +@@ -12,8 +12,6 @@ config ARC
> > + 	select BUILDTIME_EXTABLE_SORT
> > + 	select COMMON_CLK
> > + 	select CLONE_BACKWARDS
> > +-	# ARC Busybox based initramfs absolutely relies on DEVTMPFS for /dev
> > +-	select DEVTMPFS if !INITRAMFS_SOURCE=""
> > + 	select GENERIC_ATOMIC64
> > + 	select GENERIC_CLOCKEVENTS
> > + 	select GENERIC_FIND_FIRST_BIT
> Could you please send this one upstream as well?

Adding Vineet so he may comment if that change makes sense upstream.

-Alexey

[PATCH 2/2] arc770: move arc patches to taregt/linux/generic

[Vineet Gupta]

+CC Thomas, Peter

On Friday 15 January 2016 04:19 PM, Alexey Brodkin wrote:
> Hi Felix,
>
> On Fri, 2016-01-15@11:45 +0100, Felix Fietkau wrote:
>> On 2016-01-15 00:12, Alexey Brodkin wrote:
>>> Given those patches are relevant to any ARC platform and even
>>> ISA version it makes perfect sense for patches to exist
>>> in one place instead of being duplicated for each new ARC-based ASIC.
>>>
>>> Note this is a prerequisite for upstreaming of ARC HS38 support in
>>> OpenWRT.
>>>
>>> Signed-off-by: Alexey Brodkin <abrodkin at synopsys.com>
>>> Cc: Felix Fietkau <nbd at openwrt.org>
>>> Cc: Jo-Philipp Wich <jow at openwrt.org>
>>> Cc: Jonas Gorski <jogo at openwrt.org>
>>> --- /dev/null
>>> +++ b/target/linux/generic/patches-4.4/331-arc-remove-dependency-on-DEVTMPFS.patch
>>> @@ -0,0 +1,36 @@
>>> +From adfbf9e6cad93281cffceab078e7f6f2a8e094f9 Mon Sep 17 00:00:00 2001
>>> +From: Alexey Brodkin <abrodkin at synopsys.com>
>>> +Date: Thu, 13 Aug 2015 01:56:02 +0300
>>> +Subject: [PATCH 1/2] openwrt: arc - remove dependency on DEVTMPFS
>>> +
>>> +OpenWRT builds initramfs so that it doesn't require DEVTMPFS so dropping
>>> +this dependency. 

Really ? AFAIKR (circa 2012) DEVTMPFS was *needed* for dynamic device nodes and
that included the common case of initramfs NOT having static device nodes.

So back then I added the code in question to kernel Kconfig because initramfs was
my primary workflow and occassionally I would fail to include DEVTMPFS causing
userspace boot to go bonkers (FWIW I was using the buildroot trick of a pre-init
script which would automount devtmpfs before exec'ing the real init)

Now arguably I can add DEVTMPFS to defconfigs, but then we don't need the kconfig
dependency framework at all.

Another idea is to add DEVTMPFS unconditionally to Kconfig, but I fail to remember
why I didn't do it at the time. Does anyone know if it interferes with real rootfs
backed by real devices ?

>>> That helps to escape 2 separate kernel rebuilds with
>>> +and without initramfs.
>>> +
>>> +2 builds happen because OpenWRT first builds kernel and later modules.
>>> +When building entire kernel with simple "make" INITRAMFS sets to a real
>>> +value and so was triggering DEVTMPFS selection. Then when building only
>>> +modules with "make modules" command INITRAMFS is zeroed and so kernel
>>> +config was changing that lead to full kernel rebuild.

Just curious, why does make module zero out INITRAMFS - to save a bit of build time ?

-Vineet

>>> +
>>> +Signed-off-by: Alexey Brodkin <abrodkin at synopsys.com>
>>> +---
>>> + arch/arc/Kconfig | 2 --
>>> + 1 file changed, 2 deletions(-)
>>> +
>>> +diff --git a/arch/arc/Kconfig b/arch/arc/Kconfig
>>> +index 6312f60..a95bab3 100644
>>> +--- a/arch/arc/Kconfig
>>> ++++ b/arch/arc/Kconfig
>>> +@@ -12,8 +12,6 @@ config ARC
>>> + 	select BUILDTIME_EXTABLE_SORT
>>> + 	select COMMON_CLK
>>> + 	select CLONE_BACKWARDS
>>> +-	# ARC Busybox based initramfs absolutely relies on DEVTMPFS for /dev
>>> +-	select DEVTMPFS if !INITRAMFS_SOURCE=""
>>> + 	select GENERIC_ATOMIC64
>>> + 	select GENERIC_CLOCKEVENTS
>>> + 	select GENERIC_FIND_FIRST_BIT
>> Could you please send this one upstream as well?
> Adding Vineet so he may comment if that change makes sense upstream.
>
> -Alexey

[PATCH 2/2] arc770: move arc patches to taregt/linux/generic

[Felix Fietkau]

On 2016-01-15 13:25, Vineet Gupta wrote:
> +CC Thomas, Peter
> 
> On Friday 15 January 2016 04:19 PM, Alexey Brodkin wrote:
>> Hi Felix,
>>
>> On Fri, 2016-01-15@11:45 +0100, Felix Fietkau wrote:
>>> On 2016-01-15 00:12, Alexey Brodkin wrote:
>>>> Given those patches are relevant to any ARC platform and even
>>>> ISA version it makes perfect sense for patches to exist
>>>> in one place instead of being duplicated for each new ARC-based ASIC.
>>>>
>>>> Note this is a prerequisite for upstreaming of ARC HS38 support in
>>>> OpenWRT.
>>>>
>>>> Signed-off-by: Alexey Brodkin <abrodkin at synopsys.com>
>>>> Cc: Felix Fietkau <nbd at openwrt.org>
>>>> Cc: Jo-Philipp Wich <jow at openwrt.org>
>>>> Cc: Jonas Gorski <jogo at openwrt.org>
>>>> --- /dev/null
>>>> +++ b/target/linux/generic/patches-4.4/331-arc-remove-dependency-on-DEVTMPFS.patch
>>>> @@ -0,0 +1,36 @@
>>>> +From adfbf9e6cad93281cffceab078e7f6f2a8e094f9 Mon Sep 17 00:00:00 2001
>>>> +From: Alexey Brodkin <abrodkin at synopsys.com>
>>>> +Date: Thu, 13 Aug 2015 01:56:02 +0300
>>>> +Subject: [PATCH 1/2] openwrt: arc - remove dependency on DEVTMPFS
>>>> +
>>>> +OpenWRT builds initramfs so that it doesn't require DEVTMPFS so dropping
>>>> +this dependency. 
> 
> Really ? AFAIKR (circa 2012) DEVTMPFS was *needed* for dynamic device nodes and
> that included the common case of initramfs NOT having static device nodes.
> 
> So back then I added the code in question to kernel Kconfig because initramfs was
> my primary workflow and occassionally I would fail to include DEVTMPFS causing
> userspace boot to go bonkers (FWIW I was using the buildroot trick of a pre-init
> script which would automount devtmpfs before exec'ing the real init)
> 
> Now arguably I can add DEVTMPFS to defconfigs, but then we don't need the kconfig
> dependency framework at all.
> 
> Another idea is to add DEVTMPFS unconditionally to Kconfig, but I fail to remember
> why I didn't do it at the time. Does anyone know if it interferes with real rootfs
> backed by real devices ?
OpenWrt works just fine without DEVTMPFS - doesn't matter if initramfs
is enabled or not. I also don't think there's anything architecture
specific about requiring DEVTMPFS or not, so I'd say it doesn't make any
sense to select it from arch/arc/Kconfig.

>>>> That helps to escape 2 separate kernel rebuilds with
>>>> +and without initramfs.
>>>> +
>>>> +2 builds happen because OpenWRT first builds kernel and later modules.
>>>> +When building entire kernel with simple "make" INITRAMFS sets to a real
>>>> +value and so was triggering DEVTMPFS selection. Then when building only
>>>> +modules with "make modules" command INITRAMFS is zeroed and so kernel
>>>> +config was changing that lead to full kernel rebuild.
> 
> Just curious, why does make module zero out INITRAMFS - to save a bit of build time ?
OpenWrt builds both initramfs and non-initramfs images, so it resets the
INITRAMFS config item when it generates the kernel config for the first
time.

- Felix

[PATCH 2/2] arc770: move arc patches to taregt/linux/generic

[Vineet Gupta]

On Friday 15 January 2016 06:29 PM, Felix Fietkau wrote:
>>>>> +Subject: [PATCH 1/2] openwrt: arc - remove dependency on DEVTMPFS
>>>>> >>>> +
>>>>> >>>> +OpenWRT builds initramfs so that it doesn't require DEVTMPFS so dropping
>>>>> >>>> +this dependency. 
>> > 
>> > Really ? AFAIKR (circa 2012) DEVTMPFS was *needed* for dynamic device nodes and
>> > that included the common case of initramfs NOT having static device nodes.
>> > 
>> > So back then I added the code in question to kernel Kconfig because initramfs was
>> > my primary workflow and occassionally I would fail to include DEVTMPFS causing
>> > userspace boot to go bonkers (FWIW I was using the buildroot trick of a pre-init
>> > script which would automount devtmpfs before exec'ing the real init)
>> > 
>> > Now arguably I can add DEVTMPFS to defconfigs, but then we don't need the kconfig
>> > dependency framework at all.
>> > 
>> > Another idea is to add DEVTMPFS unconditionally to Kconfig, but I fail to remember
>> > why I didn't do it at the time. Does anyone know if it interferes with real rootfs
>> > backed by real devices ?
> OpenWrt works just fine without DEVTMPFS - doesn't matter if initramfs
> is enabled or not.

The discussion is about adding a patch to up upstream ARC kernel, not specific to
openwrt.
BTW if openwrt builds for initramfs, it has to enable DEVTMPFS under the hood.
Perhaps there are dependencies in openwrt build system which take care of that
already - o/w it just won't work (assuming dynamic dev nodes).

>  I also don't think there's anything architecture
> specific about requiring DEVTMPFS or not, so I'd say it doesn't make any
> sense to select it from arch/arc/Kconfig.

You are indeed right - Kconfig is not the "ideal" place to add this (it's not like
we support PCI or not) and it best belongs to *defconfigs.

-Vineet

[PATCH 2/2] arc770: move arc patches to taregt/linux/generic

[Felix Fietkau]

On 2016-01-15 14:17, Vineet Gupta wrote:
> On Friday 15 January 2016 06:29 PM, Felix Fietkau wrote:
>>>>>> +Subject: [PATCH 1/2] openwrt: arc - remove dependency on DEVTMPFS
>>>>>> >>>> +
>>>>>> >>>> +OpenWRT builds initramfs so that it doesn't require DEVTMPFS so dropping
>>>>>> >>>> +this dependency. 
>>> > 
>>> > Really ? AFAIKR (circa 2012) DEVTMPFS was *needed* for dynamic device nodes and
>>> > that included the common case of initramfs NOT having static device nodes.
>>> > 
>>> > So back then I added the code in question to kernel Kconfig because initramfs was
>>> > my primary workflow and occassionally I would fail to include DEVTMPFS causing
>>> > userspace boot to go bonkers (FWIW I was using the buildroot trick of a pre-init
>>> > script which would automount devtmpfs before exec'ing the real init)
>>> > 
>>> > Now arguably I can add DEVTMPFS to defconfigs, but then we don't need the kconfig
>>> > dependency framework at all.
>>> > 
>>> > Another idea is to add DEVTMPFS unconditionally to Kconfig, but I fail to remember
>>> > why I didn't do it at the time. Does anyone know if it interferes with real rootfs
>>> > backed by real devices ?
>> OpenWrt works just fine without DEVTMPFS - doesn't matter if initramfs
>> is enabled or not.
> 
> The discussion is about adding a patch to up upstream ARC kernel, not specific to
> openwrt.
Right. I belive that the upstream kernel should not arbitrarily force
DEVTMPFS support for some architectures, as long as there are user space
implementations (such as OpenWrt) that can do without it.

> BTW if openwrt builds for initramfs, it has to enable DEVTMPFS under the hood.
> Perhaps there are dependencies in openwrt build system which take care of that
> already - o/w it just won't work (assuming dynamic dev nodes).
Incorrect. OpenWrt does not use DEVTMPFS, it does not even get compiled
into the image. We would like to keep it that way.
Our user space takes care of creating all required device nodes very
early during boot.

- Felix

[PATCH 2/2] arc770: move arc patches to taregt/linux/generic

[Peter Korsgaard]

>>>>> "Felix" == Felix Fietkau <nbd at openwrt.org> writes:

 >>> OpenWrt works just fine without DEVTMPFS - doesn't matter if initramfs
 >>> is enabled or not.
 >> 
 >> The discussion is about adding a patch to up upstream ARC kernel, not specific to
 >> openwrt.
 > Right. I belive that the upstream kernel should not arbitrarily force
 > DEVTMPFS support for some architectures, as long as there are user space
 > implementations (such as OpenWrt) that can do without it.

Agreed, only the options absolutely needed should be forced on.

 >> BTW if openwrt builds for initramfs, it has to enable DEVTMPFS under the hood.
 >> Perhaps there are dependencies in openwrt build system which take care of that
 >> already - o/w it just won't work (assuming dynamic dev nodes).
 > Incorrect. OpenWrt does not use DEVTMPFS, it does not even get compiled
 > into the image. We would like to keep it that way.
 > Our user space takes care of creating all required device nodes very
 > early during boot.

Out of interest, why is that? Devtmpfs got added 7 years ago (2.6.32) -
And is easy to backport if really needed, is easier and more flexible
than a bunch of static mknods, and probably smaller as well.

We changed to devtmpfs by default in Buildroot quite some time ago, and
I'm pretty happy with it.

-- 
Venlig hilsen,
Peter Korsgaard 

[PATCH 2/2] arc770: move arc patches to taregt/linux/generic

[Felix Fietkau]

On 2016-01-16 11:35, Peter Korsgaard wrote:
>>>>>> "Felix" == Felix Fietkau <nbd at openwrt.org> writes:
> 
>  >>> OpenWrt works just fine without DEVTMPFS - doesn't matter if initramfs
>  >>> is enabled or not.
>  >> 
>  >> The discussion is about adding a patch to up upstream ARC kernel, not specific to
>  >> openwrt.
>  > Right. I belive that the upstream kernel should not arbitrarily force
>  > DEVTMPFS support for some architectures, as long as there are user space
>  > implementations (such as OpenWrt) that can do without it.
> 
> Agreed, only the options absolutely needed should be forced on.
> 
>  >> BTW if openwrt builds for initramfs, it has to enable DEVTMPFS under the hood.
>  >> Perhaps there are dependencies in openwrt build system which take care of that
>  >> already - o/w it just won't work (assuming dynamic dev nodes).
>  > Incorrect. OpenWrt does not use DEVTMPFS, it does not even get compiled
>  > into the image. We would like to keep it that way.
>  > Our user space takes care of creating all required device nodes very
>  > early during boot.
> 
> Out of interest, why is that? Devtmpfs got added 7 years ago (2.6.32) -
> And is easy to backport if really needed, is easier and more flexible
> than a bunch of static mknods, and probably smaller as well.
We don't need to backport anything. Our oldest kernel is 3.18, and we're
going to move everything to 4.4 soon ;)

> We changed to devtmpfs by default in Buildroot quite some time ago, and
> I'm pretty happy with it.
We need to have dynamically created device nodes anyway - for managing
permissions, being able to change names, etc. Because of that, devtmpfs
is not enough to provide a full /dev. Since it's not enough, and
creating the initial device nodes from our custom init is easy, we see
little value in keeping it. So we got rid of the extra bloat :)

- Felix

[PATCH 2/2] arc770: move arc patches to taregt/linux/generic

[Peter Korsgaard]

>>>>> "Felix" == Felix Fietkau <nbd at openwrt.org> writes:

Hi,

>> Out of interest, why is that? Devtmpfs got added 7 years ago (2.6.32) -
 >> And is easy to backport if really needed, is easier and more flexible
 >> than a bunch of static mknods, and probably smaller as well.
 > We don't need to backport anything. Our oldest kernel is 3.18, and we're
 > going to move everything to 4.4 soon ;)

Good!

 >> We changed to devtmpfs by default in Buildroot quite some time ago, and
 >> I'm pretty happy with it.
 > We need to have dynamically created device nodes anyway - for managing
 > permissions, being able to change names, etc. Because of that, devtmpfs
 > is not enough to provide a full /dev. Since it's not enough, and
 > creating the initial device nodes from our custom init is easy, we see
 > little value in keeping it. So we got rid of the extra bloat :)

Heh, "bloat":

arm-none-eabi-size drivers/base/devtmpfs.o
   text    data     bss     dec     hex filename
   1568      64       4    1636     664 drivers/base/devtmpfs.o

Compared to the extra inodes and/or the busybox mknod applet + script,
it isn't too bad.

In Buildroot we support pure devtmpfs, mdev or udev (both with devtmpfs)
or static /dev for legacy setups.

-- 
Venlig hilsen,
Peter Korsgaard 

[PATCH 2/2] arc770: move arc patches to taregt/linux/generic

[Felix Fietkau]

On 2016-01-16 12:00, Peter Korsgaard wrote:
>  >> We changed to devtmpfs by default in Buildroot quite some time ago, and
>  >> I'm pretty happy with it.
>  > We need to have dynamically created device nodes anyway - for managing
>  > permissions, being able to change names, etc. Because of that, devtmpfs
>  > is not enough to provide a full /dev. Since it's not enough, and
>  > creating the initial device nodes from our custom init is easy, we see
>  > little value in keeping it. So we got rid of the extra bloat :)
> 
> Heh, "bloat":
> 
> arm-none-eabi-size drivers/base/devtmpfs.o
>    text    data     bss     dec     hex filename
>    1568      64       4    1636     664 drivers/base/devtmpfs.o
> 
> Compared to the extra inodes and/or the busybox mknod applet + script,
> it isn't too bad.
I know it's not much, but we do like to disable anything that's
completely useless for our purposes. :)
Our code doesn't use busybox mknod. The stuff that creates the device
nodes is written in C.

> In Buildroot we support pure devtmpfs, mdev or udev (both with devtmpfs)
> or static /dev for legacy setups.
Yeah, for Buildroot devtmpfs makes sense. We don't use or support mdev,
udev or static /dev. We also got rid of hotplug2 a while back ;)

- Felix

[PATCH 2/2] arc770: move arc patches to taregt/linux/generic

[Vineet Gupta]

On Friday 15 January 2016 06:47 PM, Vineet Gupta wrote:
>>  I also don't think there's anything architecture
>> > specific about requiring DEVTMPFS or not, so I'd say it doesn't make any
>> > sense to select it from arch/arc/Kconfig.
> You are indeed right - Kconfig is not the "ideal" place to add this (it's not like
> we support PCI or not) and it best belongs to *defconfigs.

Alexey do u mind spinning a patchset which
1. Adds DEVTMFS to our defconfigs where we need them
2. Removes the select for same from Kconfig

-Vineet