lkml.org 
[lkml]   [2017]   [Sep]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v5 1/3] mfd: Add support for Cherry Trail Dollar Cove TI PMIC
On Wed, 06 Sep 2017, Takashi Iwai wrote:

> On Wed, 06 Sep 2017 11:05:04 +0200,
> Lee Jones wrote:
> >
> > On Wed, 06 Sep 2017, Takashi Iwai wrote:
> >
> > > On Wed, 06 Sep 2017 09:54:44 +0200,
> > > Lee Jones wrote:
> > > >
> > > > On Tue, 05 Sep 2017, Takashi Iwai wrote:
> > > >
> > > > > On Tue, 05 Sep 2017 10:53:41 +0200,
> > > > > Lee Jones wrote:
> > > > > >
> > > > > > On Tue, 05 Sep 2017, Takashi Iwai wrote:
> > > > > >
> > > > > > > On Tue, 05 Sep 2017 10:10:49 +0200,
> > > > > > > Lee Jones wrote:
> > > > > > > >
> > > > > > > > On Tue, 05 Sep 2017, Takashi Iwai wrote:
> > > > > > > >
> > > > > > > > > On Tue, 05 Sep 2017 09:24:51 +0200,
> > > > > > > > > Lee Jones wrote:
> > > > > > > > > >
> > > > > > > > > > On Mon, 04 Sep 2017, Takashi Iwai wrote:
> > > > > > > > > >
> > > > > > > > > > > This patch adds the MFD driver for Dollar Cove (TI version) PMIC with
> > > > > > > > > > > ACPI INT33F5 that is found on some Intel Cherry Trail devices.
> > > > > > > > > > > The driver is based on the original work by Intel, found at:
> > > > > > > > > > > https://github.com/01org/ProductionKernelQuilts
> > > > > > > > > > >
> > > > > > > > > > > This is a minimal version for adding the basic resources. Currently,
> > > > > > > > > > > only ACPI PMIC opregion and the external power-button are used.
> > > > > > > > > > >
> > > > > > > > > > > Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=193891
> > > > > > > > > > > Reviewed-by: Mika Westerberg <mika.westerberg@linux.intel.com>
> > > > > > > > > > > Reviewed-by: Andy Shevchenko <andy.shevchenko@gmail.com>
> > > > > > > > > > > Signed-off-by: Takashi Iwai <tiwai@suse.de>
> > > > > > > > > > > ---
> > > > > > > > > > > v4->v5:
> > > > > > > > > > > * Minor coding-style fixes suggested by Lee
> > > > > > > > > > > * Put GPL text
> > > > > > > > > > > v3->v4:
> > > > > > > > > > > * no change for this patch
> > > > > > > > > > > v2->v3:
> > > > > > > > > > > * Rename dc_ti with chtdc_ti in all places
> > > > > > > > > > > * Driver/kconfig renames accordingly
> > > > > > > > > > > * Added acks by Andy and Mika
> > > > > > > > > > > v1->v2:
> > > > > > > > > > > * Minor cleanups as suggested by Andy
> > > > > > > > > > >
> > > > > > > > > > > drivers/mfd/Kconfig | 13 +++
> > > > > > > > > > > drivers/mfd/Makefile | 1 +
> > > > > > > > > > > drivers/mfd/intel_soc_pmic_chtdc_ti.c | 184 ++++++++++++++++++++++++++++++++++
> > > > > > > > > > > 3 files changed, 198 insertions(+)
> > > > > > > > > > > create mode 100644 drivers/mfd/intel_soc_pmic_chtdc_ti.c
> > > > > > > > > >
> > > > > > > > > > For my own reference:
> > > > > > > > > > Acked-for-MFD-by: Lee Jones <lee.jones@linaro.org>
> > > > > > > > >
> > > > > > > > > Thanks!
> > > > > > > > >
> > > > > > > > > Now the question is how to deal with these. It's no critical things,
> > > > > > > > > so I'm OK to postpone for 4.15. OTOH, it's really a new
> > > > > > > > > device-specific stuff, thus it can't break anything else, and it'd be
> > > > > > > > > fairly safe to add it for 4.14 although it's at a bit late stage.
> > > > > > > >
> > > > > > > > Yes, you are over 2 weeks late for v4.14. It will have to be v4.15.
> > > > > > >
> > > > > > > OK, I'll ring your bells again once when 4.15 development is opened.
> > > > > > >
> > > > > > >
> > > > > > > > > IMO, it'd be great if you can carry all stuff through MFD tree; or
> > > > > > > > > create an immutable branch (again). But how to handle it, when to do
> > > > > > > > > it, It's all up to you guys.
> > > > > > > >
> > > > > > > > If there aren't any build dependencies between the patches, each of
> > > > > > > > the patches should be applied through their own trees. What are the
> > > > > > > > build-time dependencies? Are there any?
> > > > > > >
> > > > > > > No, there is no strict build-time dependency. It's just that I don't
> > > > > > > see it nice to have a commit for a dead code, partly for testing
> > > > > > > purpose and partly for code consistency. But if this makes
> > > > > > > maintenance easier, I'm happy with that, too, of course.
> > > > > >
> > > > > > There won't be any dead code. All of the subsystem trees are pulled
> > > > > > into -next [0] where the build bots can operate on the patches as a
> > > > > > whole.
> > > > >
> > > > > But the merge order isn't guaranteed, i.e. at the commit of other tree
> > > > > for this new stuff, it's a dead code without merging the MFD stuff
> > > > > beforehand. e.g. Imagine to perform the git bisection. It's not
> > > > > about the whole tree, but about the each commit.
> > > >
> > > > Only *building* is relevant for bisection until the whole feature
> > > > lands.
> > >
> > > Why only building?
> > >
> > > When merging through several tress, commits for the same series are
> > > scattered completely although they are softly tied. This sucks when
> > > you perform git bisection, e.g. if you have an issue in the middle of
> > > the patch series. It still works, but it jumps unnecessarily too far
> > > away and back before reaching to the point, and kconfig appears /
> > > disappears inconsistently (the dependent kconfig gone in the middle).
> > > And, this is about the release kernel (4.15 or whatever).
> >
> > Think about how bisection works. You state a good commit and a bad
> > one. The good commit will be when the feature last worked, which will
> > not be until the feature has fully landed. Bisect will not check any
> > point prior to this date.
> >
> > If there aren't any build deps, each Maintainer will apply patches
> > into their own tree. These will be merged together in -next where
> > they can be tested, both manually and by the 0-days. Once the merge
> > window is opened all patches will be sucked into -rc1. If the feature
> > works here, then it you could use -rc1 as your 'good' commit. If it
> > doesn't, then this could indicate a merge error or a missing piece of
> > the set, either way bisect wouldn't help you.
>
> Not really.
>
> First of all, most of user start testing from the release kernel, so
> you can't trust that RC covered all test cases. (Who can blame users
> who didn't use / test RC?)

That's fine. We are not talking about spreading the merge of a
patch-set over different releases, or even release candidates. All
non-bugfix patches should be in by -rc1.

> Second, you ignore the fact that the development continues after
> merging *this* patchset. What if a breakage is introduced after this
> patch? (See below)
>
> They often need a full bisection between the previous release and the
> current release.

You cannot bisect a specific function back before it was merged. It's
impossible.

P1 ---> P4 ---> P3 ---> P2
^
Bisect only starts working here. Prior to this point the feature
doesn't build at all. If it builds, but breaks, then that is a build
dependency and is a different use-case to what we are discussing here.

> > > Basically, my complaint here comes with my user's hat on. It *is*
> > > indeed worse than a straight application of patches in some levels.
> > > It's unavoidable if you do in that way.
> >
> > I disagree. I user wouldn't set the 'good' commit at any point prior
> > to the feature working at least once.
>
> Again no, not all users do test at the same time. A device driver
> may support multiple devices / platforms, and it might be that the bug
> manifests itself only on a certain system that no one has tested
> beforehand; it's a typical case we often see after the releases.

If it is possible for a patch-set to break functionality, then again,
this is a dependency and needs to be handled differently.

That is not the case with your set.

> > This will not happen if any
> > parts were missing. The order in which the pieces are applied is
> > irrelevant if there aren't any build deps between them. If you bisect
> > between them, then the driver simply will not build. No problem.
>
> The bisection is required not only for build errors but also for
> functional tests. Sometimes it's the only way to spot the culprit of
> a functional regression.
>
> Imagine the following case: both MFD and platform drivers are merged
> into individual trees separately. And, some change in platform driver
> after this patch series broke some functionality mistakenly.
>
> Now, suppose that the platform tree is merged to Linus tree before MFD
> tree, that is, the situation is like:
>
> MFD (commit A1) -- ... -------------------------+
> platform (commit B1) -> broken change (B2) -+ |
> v v
> --M1--M2--> Linus RC1
>
> git-bisection can inspect M2, which shows already broken. Then it
> inspects M1, but you can't evaluate it because the target platform
> driver can't be built without A1. At this point, you're stuck.
> But in this case, B1 is correct and B2 is the culprit. How can you
> spot it out?
>
> OTOH, if the merge history honors the functional dependency, you can
> bisect properly.

You wouldn't use bisect for that use-case. We only use bisect for
features that have a working commit in Mainline. The use-case you
mention would be pretty trivial to debug without bisect.

When we tout "history must be bisectable", we mean that it must remain
possible to bisect existing functionality. Bisecting functionality
which is yet to fully land upstream is foolish.

New drivers almost never (if ever) go in as a fully working piece.
The registration parts (DT, Platform, etc) go in via one (or several)
tree(s) and the driver parts are normally split up into subsystems and
go in via their own repos. If there are *build*, *API* or *merge*
time dependencies, then we can use immutable branches and merge via
a single tree. Failing that, we split them.

--
Lee Jones
Linaro STMicroelectronics Landing Team Lead
Linaro.org │ Open source software for ARM SoCs
Follow Linaro: Facebook | Twitter | Blog

\
 
 \ /
  Last update: 2017-09-06 12:41    [W:0.110 / U:8.028 seconds]
©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site