blob: 1f5ab49c48a480d20f01b4269867f7c7a6451764 [file] [log] [blame]
.. SPDX-License-Identifier: (GPL-2.0+ OR CC-BY-4.0)
.. See the bottom of this file for additional redistribution information.
Handling regressions
++++++++++++++++++++
*We don't cause regressions* -- this document describes what this "first rule of
Linux kernel development" means in practice for developers. It complements
Documentation/admin-guide/reporting-regressions.rst, which covers the topic from a
user's point of view; if you never read that text, go and at least skim over it
before continuing here.
The important bits (aka "The TL;DR")
====================================
#. Ensure subscribers of the `regression mailing list <https://lore.kernel.org/regressions/>`_
(regressions@lists.linux.dev) quickly become aware of any new regression
report:
* When receiving a mailed report that did not CC the list, bring it into the
loop by immediately sending at least a brief "Reply-all" with the list
CCed.
* Forward or bounce any reports submitted in bug trackers to the list.
#. Make the Linux kernel regression tracking bot "regzbot" track the issue (this
is optional, but recommended):
* For mailed reports, check if the reporter included a line like ``#regzbot
introduced: v5.13..v5.14-rc1``. If not, send a reply (with the regressions
list in CC) containing a paragraph like the following, which tells regzbot
when the issue started to happen::
#regzbot ^introduced: 1f2e3d4c5b6a
* When forwarding reports from a bug tracker to the regressions list (see
above), include a paragraph like the following::
#regzbot introduced: v5.13..v5.14-rc1
#regzbot from: Some N. Ice Human <some.human@example.com>
#regzbot monitor: http://some.bugtracker.example.com/ticket?id=123456789
#. When submitting fixes for regressions, add "Closes:" tags to the patch
description pointing to all places where the issue was reported, as
mandated by Documentation/process/submitting-patches.rst and
:ref:`Documentation/process/5.Posting.rst <development_posting>`. If you are
only fixing part of the issue that caused the regression, you may use
"Link:" tags instead. regzbot currently makes no distinction between the
two.
#. Try to fix regressions quickly once the culprit has been identified; fixes
for most regressions should be merged within two weeks, but some need to be
resolved within two or three days.
All the details on Linux kernel regressions relevant for developers
===================================================================
The important basics in more detail
-----------------------------------
What to do when receiving regression reports
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Ensure the Linux kernel's regression tracker and others subscribers of the
`regression mailing list <https://lore.kernel.org/regressions/>`_
(regressions@lists.linux.dev) become aware of any newly reported regression:
* When you receive a report by mail that did not CC the list, immediately bring
it into the loop by sending at least a brief "Reply-all" with the list CCed;
try to ensure it gets CCed again in case you reply to a reply that omitted
the list.
* If a report submitted in a bug tracker hits your Inbox, forward or bounce it
to the list. Consider checking the list archives beforehand, if the reporter
already forwarded the report as instructed by
Documentation/admin-guide/reporting-issues.rst.
When doing either, consider making the Linux kernel regression tracking bot
"regzbot" immediately start tracking the issue:
* For mailed reports, check if the reporter included a "regzbot command" like
``#regzbot introduced: 1f2e3d4c5b6a``. If not, send a reply (with the
regressions list in CC) with a paragraph like the following:::
#regzbot ^introduced: v5.13..v5.14-rc1
This tells regzbot the version range in which the issue started to happen;
you can specify a range using commit-ids as well or state a single commit-id
in case the reporter bisected the culprit.
Note the caret (^) before the "introduced": it tells regzbot to treat the
parent mail (the one you reply to) as the initial report for the regression
you want to see tracked; that's important, as regzbot will later look out
for patches with "Closes:" tags pointing to the report in the archives on
lore.kernel.org.
* When forwarding a regression reported to a bug tracker, include a paragraph
with these regzbot commands::
#regzbot introduced: 1f2e3d4c5b6a
#regzbot from: Some N. Ice Human <some.human@example.com>
#regzbot monitor: http://some.bugtracker.example.com/ticket?id=123456789
Regzbot will then automatically associate patches with the report that
contain "Closes:" tags pointing to your mail or the mentioned ticket.
What's important when fixing regressions
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
You don't need to do anything special when submitting fixes for regression, just
remember to do what Documentation/process/submitting-patches.rst,
:ref:`Documentation/process/5.Posting.rst <development_posting>`, and
Documentation/process/stable-kernel-rules.rst already explain in more detail:
* Point to all places where the issue was reported using "Closes:" tags::
Closes: https://lore.kernel.org/r/30th.anniversary.repost@klaava.Helsinki.FI/
Closes: https://bugzilla.kernel.org/show_bug.cgi?id=1234567890
If you are only fixing part of the issue, you may use "Link:" instead as
described in the first document mentioned above. regzbot currently treats
both of these equivalently and considers the linked reports as resolved.
* Add a "Fixes:" tag to specify the commit causing the regression.
* If the culprit was merged in an earlier development cycle, explicitly mark
the fix for backporting using the ``Cc: stable@vger.kernel.org`` tag.
All this is expected from you and important when it comes to regression, as
these tags are of great value for everyone (you included) that might be looking
into the issue weeks, months, or years later. These tags are also crucial for
tools and scripts used by other kernel developers or Linux distributions; one of
these tools is regzbot, which heavily relies on the "Closes:" tags to associate
reports for regression with changes resolving them.
Expectations and best practices for fixing regressions
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
As a Linux kernel developer, you are expected to give your best to prevent
situations where a regression caused by a recent change of yours leaves users
only these options:
* Run a kernel with a regression that impacts usage.
* Switch to an older or newer kernel series.
* Continue running an outdated and thus potentially insecure kernel for more
than three weeks after the regression's culprit was identified. Ideally it
should be less than two. And it ought to be just a few days, if the issue is
severe or affects many users -- either in general or in prevalent
environments.
How to realize that in practice depends on various factors. Use the following
rules of thumb as a guide.
In general:
* Prioritize work on regressions over all other Linux kernel work, unless the
latter concerns a severe issue (e.g. acute security vulnerability, data loss,
bricked hardware, ...).
* Expedite fixing mainline regressions that recently made it into a proper
mainline, stable, or longterm release (either directly or via backport).
* Do not consider regressions from the current cycle as something that can wait
till the end of the cycle, as the issue might discourage or prevent users and
CI systems from testing mainline now or generally.
* Work with the required care to avoid additional or bigger damage, even if
resolving an issue then might take longer than outlined below.
On timing once the culprit of a regression is known:
* Aim to mainline a fix within two or three days, if the issue is severe or
bothering many users -- either in general or in prevalent conditions like a
particular hardware environment, distribution, or stable/longterm series.
* Aim to mainline a fix by Sunday after the next, if the culprit made it
into a recent mainline, stable, or longterm release (either directly or via
backport); if the culprit became known early during a week and is simple to
resolve, try to mainline the fix within the same week.
* For other regressions, aim to mainline fixes before the hindmost Sunday
within the next three weeks. One or two Sundays later are acceptable, if the
regression is something people can live with easily for a while -- like a
mild performance regression.
* It's strongly discouraged to delay mainlining regression fixes till the next
merge window, except when the fix is extraordinarily risky or when the
culprit was mainlined more than a year ago.
On procedure:
* Always consider reverting the culprit, as it's often the quickest and least
dangerous way to fix a regression. Don't worry about mainlining a fixed
variant later: that should be straight-forward, as most of the code went
through review once already.
* Try to resolve any regressions introduced in mainline during the past
twelve months before the current development cycle ends: Linus wants such
regressions to be handled like those from the current cycle, unless fixing
bears unusual risks.
* Consider CCing Linus on discussions or patch review, if a regression seems
tangly. Do the same in precarious or urgent cases -- especially if the
subsystem maintainer might be unavailable. Also CC the stable team, when you
know such a regression made it into a mainline, stable, or longterm release.
* For urgent regressions, consider asking Linus to pick up the fix straight
from the mailing list: he is totally fine with that for uncontroversial
fixes. Ideally though such requests should happen in accordance with the
subsystem maintainers or come directly from them.
* In case you are unsure if a fix is worth the risk applying just days before
a new mainline release, send Linus a mail with the usual lists and people in
CC; in it, summarize the situation while asking him to consider picking up
the fix straight from the list. He then himself can make the call and when
needed even postpone the release. Such requests again should ideally happen
in accordance with the subsystem maintainers or come directly from them.
Regarding stable and longterm kernels:
* You are free to leave regressions to the stable team, if they at no point in
time occurred with mainline or were fixed there already.
* If a regression made it into a proper mainline release during the past
twelve months, ensure to tag the fix with "Cc: stable@vger.kernel.org", as a
"Fixes:" tag alone does not guarantee a backport. Please add the same tag,
in case you know the culprit was backported to stable or longterm kernels.
* When receiving reports about regressions in recent stable or longterm kernel
series, please evaluate at least briefly if the issue might happen in current
mainline as well -- and if that seems likely, take hold of the report. If in
doubt, ask the reporter to check mainline.
* Whenever you want to swiftly resolve a regression that recently also made it
into a proper mainline, stable, or longterm release, fix it quickly in
mainline; when appropriate thus involve Linus to fast-track the fix (see
above). That's because the stable team normally does neither revert nor fix
any changes that cause the same problems in mainline.
* In case of urgent regression fixes you might want to ensure prompt
backporting by dropping the stable team a note once the fix was mainlined;
this is especially advisable during merge windows and shortly thereafter, as
the fix otherwise might land at the end of a huge patch queue.
On patch flow:
* Developers, when trying to reach the time periods mentioned above, remember
to account for the time it takes to get fixes tested, reviewed, and merged by
Linus, ideally with them being in linux-next at least briefly. Hence, if a
fix is urgent, make it obvious to ensure others handle it appropriately.
* Reviewers, you are kindly asked to assist developers in reaching the time
periods mentioned above by reviewing regression fixes in a timely manner.
* Subsystem maintainers, you likewise are encouraged to expedite the handling
of regression fixes. Thus evaluate if skipping linux-next is an option for
the particular fix. Also consider sending git pull requests more often than
usual when needed. And try to avoid holding onto regression fixes over
weekends -- especially when the fix is marked for backporting.
More aspects regarding regressions developers should be aware of
----------------------------------------------------------------
How to deal with changes where a risk of regression is known
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Evaluate how big the risk of regressions is, for example by performing a code
search in Linux distributions and Git forges. Also consider asking other
developers or projects likely to be affected to evaluate or even test the
proposed change; if problems surface, maybe some solution acceptable for all
can be found.
If the risk of regressions in the end seems to be relatively small, go ahead
with the change, but let all involved parties know about the risk. Hence, make
sure your patch description makes this aspect obvious. Once the change is
merged, tell the Linux kernel's regression tracker and the regressions mailing
list about the risk, so everyone has the change on the radar in case reports
trickle in. Depending on the risk, you also might want to ask the subsystem
maintainer to mention the issue in his mainline pull request.
What else is there to known about regressions?
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Check out Documentation/admin-guide/reporting-regressions.rst, it covers a lot
of other aspects you want might want to be aware of:
* the purpose of the "no regressions" rule
* what issues actually qualify as regression
* who's in charge for finding the root cause of a regression
* how to handle tricky situations, e.g. when a regression is caused by a
security fix or when fixing a regression might cause another one
Whom to ask for advice when it comes to regressions
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Send a mail to the regressions mailing list (regressions@lists.linux.dev) while
CCing the Linux kernel's regression tracker (regressions@leemhuis.info); if the
issue might better be dealt with in private, feel free to omit the list.
More about regression tracking and regzbot
------------------------------------------
Why the Linux kernel has a regression tracker, and why is regzbot used?
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Rules like "no regressions" need someone to ensure they are followed, otherwise
they are broken either accidentally or on purpose. History has shown this to be
true for the Linux kernel as well. That's why Thorsten Leemhuis volunteered to
keep an eye on things as the Linux kernel's regression tracker, who's
occasionally helped by other people. Neither of them are paid to do this,
that's why regression tracking is done on a best effort basis.
Earlier attempts to manually track regressions have shown it's an exhausting and
frustrating work, which is why they were abandoned after a while. To prevent
this from happening again, Thorsten developed regzbot to facilitate the work,
with the long term goal to automate regression tracking as much as possible for
everyone involved.
How does regression tracking work with regzbot?
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The bot watches for replies to reports of tracked regressions. Additionally,
it's looking out for posted or committed patches referencing such reports
with "Closes:" tags; replies to such patch postings are tracked as well.
Combined this data provides good insights into the current state of the fixing
process.
Regzbot tries to do its job with as little overhead as possible for both
reporters and developers. In fact, only reporters are burdened with an extra
duty: they need to tell regzbot about the regression report using the ``#regzbot
introduced`` command outlined above; if they don't do that, someone else can
take care of that using ``#regzbot ^introduced``.
For developers there normally is no extra work involved, they just need to make
sure to do something that was expected long before regzbot came to light: add
links to the patch description pointing to all reports about the issue fixed.
Do I have to use regzbot?
~~~~~~~~~~~~~~~~~~~~~~~~~
It's in the interest of everyone if you do, as kernel maintainers like Linus
Torvalds partly rely on regzbot's tracking in their work -- for example when
deciding to release a new version or extend the development phase. For this they
need to be aware of all unfixed regression; to do that, Linus is known to look
into the weekly reports sent by regzbot.
Do I have to tell regzbot about every regression I stumble upon?
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Ideally yes: we are all humans and easily forget problems when something more
important unexpectedly comes up -- for example a bigger problem in the Linux
kernel or something in real life that's keeping us away from keyboards for a
while. Hence, it's best to tell regzbot about every regression, except when you
immediately write a fix and commit it to a tree regularly merged to the affected
kernel series.
How to see which regressions regzbot tracks currently?
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Check `regzbot's web-interface <https://linux-regtracking.leemhuis.info/regzbot/>`_
for the latest info; alternatively, `search for the latest regression report
<https://lore.kernel.org/lkml/?q=%22Linux+regressions+report%22+f%3Aregzbot>`_,
which regzbot normally sends out once a week on Sunday evening (UTC), which is a
few hours before Linus usually publishes new (pre-)releases.
What places is regzbot monitoring?
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Regzbot is watching the most important Linux mailing lists as well as the git
repositories of linux-next, mainline, and stable/longterm.
What kind of issues are supposed to be tracked by regzbot?
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The bot is meant to track regressions, hence please don't involve regzbot for
regular issues. But it's okay for the Linux kernel's regression tracker if you
use regzbot to track severe issues, like reports about hangs, corrupted data,
or internal errors (Panic, Oops, BUG(), warning, ...).
Can I add regressions found by CI systems to regzbot's tracking?
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Feel free to do so, if the particular regression likely has impact on practical
use cases and thus might be noticed by users; hence, please don't involve
regzbot for theoretical regressions unlikely to show themselves in real world
usage.
How to interact with regzbot?
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
By using a 'regzbot command' in a direct or indirect reply to the mail with the
regression report. These commands need to be in their own paragraph (IOW: they
need to be separated from the rest of the mail using blank lines).
One such command is ``#regzbot introduced: <version or commit>``, which makes
regzbot consider your mail as a regressions report added to the tracking, as
already described above; ``#regzbot ^introduced: <version or commit>`` is another
such command, which makes regzbot consider the parent mail as a report for a
regression which it starts to track.
Once one of those two commands has been utilized, other regzbot commands can be
used in direct or indirect replies to the report. You can write them below one
of the `introduced` commands or in replies to the mail that used one of them
or itself is a reply to that mail:
* Set or update the title::
#regzbot title: foo
* Monitor a discussion or bugzilla.kernel.org ticket where additions aspects of
the issue or a fix are discussed -- for example the posting of a patch fixing
the regression::
#regzbot monitor: https://lore.kernel.org/all/30th.anniversary.repost@klaava.Helsinki.FI/
Monitoring only works for lore.kernel.org and bugzilla.kernel.org; regzbot
will consider all messages in that thread or ticket as related to the fixing
process.
* Point to a place with further details of interest, like a mailing list post
or a ticket in a bug tracker that are slightly related, but about a different
topic::
#regzbot link: https://bugzilla.kernel.org/show_bug.cgi?id=123456789
* Mark a regression as fixed by a commit that is heading upstream or already
landed::
#regzbot fix: 1f2e3d4c5d
* Mark a regression as a duplicate of another one already tracked by regzbot::
#regzbot dup-of: https://lore.kernel.org/all/30th.anniversary.repost@klaava.Helsinki.FI/
* Mark a regression as invalid::
#regzbot invalid: wasn't a regression, problem has always existed
Is there more to tell about regzbot and its commands?
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
More detailed and up-to-date information about the Linux
kernel's regression tracking bot can be found on its
`project page <https://gitlab.com/knurd42/regzbot>`_, which among others
contains a `getting started guide <https://gitlab.com/knurd42/regzbot/-/blob/main/docs/getting_started.md>`_
and `reference documentation <https://gitlab.com/knurd42/regzbot/-/blob/main/docs/reference.md>`_
which both cover more details than the above section.
Quotes from Linus about regression
----------------------------------
Find below a few real life examples of how Linus Torvalds expects regressions to
be handled:
* From `2017-10-26 (1/2)
<https://lore.kernel.org/lkml/CA+55aFwiiQYJ+YoLKCXjN_beDVfu38mg=Ggg5LFOcqHE8Qi7Zw@mail.gmail.com/>`_::
If you break existing user space setups THAT IS A REGRESSION.
It's not ok to say "but we'll fix the user space setup".
Really. NOT OK.
[...]
The first rule is:
- we don't cause regressions
and the corollary is that when regressions *do* occur, we admit to
them and fix them, instead of blaming user space.
The fact that you have apparently been denying the regression now for
three weeks means that I will revert, and I will stop pulling apparmor
requests until the people involved understand how kernel development
is done.
* From `2017-10-26 (2/2)
<https://lore.kernel.org/lkml/CA+55aFxW7NMAMvYhkvz1UPbUTUJewRt6Yb51QAx5RtrWOwjebg@mail.gmail.com/>`_::
People should basically always feel like they can update their kernel
and simply not have to worry about it.
I refuse to introduce "you can only update the kernel if you also
update that other program" kind of limitations. If the kernel used to
work for you, the rule is that it continues to work for you.
There have been exceptions, but they are few and far between, and they
generally have some major and fundamental reasons for having happened,
that were basically entirely unavoidable, and people _tried_hard_ to
avoid them. Maybe we can't practically support the hardware any more
after it is decades old and nobody uses it with modern kernels any
more. Maybe there's a serious security issue with how we did things,
and people actually depended on that fundamentally broken model. Maybe
there was some fundamental other breakage that just _had_ to have a
flag day for very core and fundamental reasons.
And notice that this is very much about *breaking* peoples environments.
Behavioral changes happen, and maybe we don't even support some
feature any more. There's a number of fields in /proc/<pid>/stat that
are printed out as zeroes, simply because they don't even *exist* in
the kernel any more, or because showing them was a mistake (typically
an information leak). But the numbers got replaced by zeroes, so that
the code that used to parse the fields still works. The user might not
see everything they used to see, and so behavior is clearly different,
but things still _work_, even if they might no longer show sensitive
(or no longer relevant) information.
But if something actually breaks, then the change must get fixed or
reverted. And it gets fixed in the *kernel*. Not by saying "well, fix
your user space then". It was a kernel change that exposed the
problem, it needs to be the kernel that corrects for it, because we
have a "upgrade in place" model. We don't have a "upgrade with new
user space".
And I seriously will refuse to take code from people who do not
understand and honor this very simple rule.
This rule is also not going to change.
And yes, I realize that the kernel is "special" in this respect. I'm
proud of it.
I have seen, and can point to, lots of projects that go "We need to
break that use case in order to make progress" or "you relied on
undocumented behavior, it sucks to be you" or "there's a better way to
do what you want to do, and you have to change to that new better
way", and I simply don't think that's acceptable outside of very early
alpha releases that have experimental users that know what they signed
up for. The kernel hasn't been in that situation for the last two
decades.
We do API breakage _inside_ the kernel all the time. We will fix
internal problems by saying "you now need to do XYZ", but then it's
about internal kernel API's, and the people who do that then also
obviously have to fix up all the in-kernel users of that API. Nobody
can say "I now broke the API you used, and now _you_ need to fix it
up". Whoever broke something gets to fix it too.
And we simply do not break user space.
* From `2020-05-21
<https://lore.kernel.org/all/CAHk-=wiVi7mSrsMP=fLXQrXK_UimybW=ziLOwSzFTtoXUacWVQ@mail.gmail.com/>`_::
The rules about regressions have never been about any kind of
documented behavior, or where the code lives.
The rules about regressions are always about "breaks user workflow".
Users are literally the _only_ thing that matters.
No amount of "you shouldn't have used this" or "that behavior was
undefined, it's your own fault your app broke" or "that used to work
simply because of a kernel bug" is at all relevant.
Now, reality is never entirely black-and-white. So we've had things
like "serious security issue" etc that just forces us to make changes
that may break user space. But even then the rule is that we don't
really have other options that would allow things to continue.
And obviously, if users take years to even notice that something
broke, or if we have sane ways to work around the breakage that
doesn't make for too much trouble for users (ie "ok, there are a
handful of users, and they can use a kernel command line to work
around it" kind of things) we've also been a bit less strict.
But no, "that was documented to be broken" (whether it's because the
code was in staging or because the man-page said something else) is
irrelevant. If staging code is so useful that people end up using it,
that means that it's basically regular kernel code with a flag saying
"please clean this up".
The other side of the coin is that people who talk about "API
stability" are entirely wrong. API's don't matter either. You can make
any changes to an API you like - as long as nobody notices.
Again, the regression rule is not about documentation, not about
API's, and not about the phase of the moon.
It's entirely about "we caused problems for user space that used to work".
* From `2017-11-05
<https://lore.kernel.org/all/CA+55aFzUvbGjD8nQ-+3oiMBx14c_6zOj2n7KLN3UsJ-qsd4Dcw@mail.gmail.com/>`_::
And our regression rule has never been "behavior doesn't change".
That would mean that we could never make any changes at all.
For example, we do things like add new error handling etc all the
time, which we then sometimes even add tests for in our kselftest
directory.
So clearly behavior changes all the time and we don't consider that a
regression per se.
The rule for a regression for the kernel is that some real user
workflow breaks. Not some test. Not a "look, I used to be able to do
X, now I can't".
* From `2018-08-03
<https://lore.kernel.org/all/CA+55aFwWZX=CXmWDTkDGb36kf12XmTehmQjbiMPCqCRG2hi9kw@mail.gmail.com/>`_::
YOU ARE MISSING THE #1 KERNEL RULE.
We do not regress, and we do not regress exactly because your are 100% wrong.
And the reason you state for your opinion is in fact exactly *WHY* you
are wrong.
Your "good reasons" are pure and utter garbage.
The whole point of "we do not regress" is so that people can upgrade
the kernel and never have to worry about it.
> Kernel had a bug which has been fixed
That is *ENTIRELY* immaterial.
Guys, whether something was buggy or not DOES NOT MATTER.
Why?
Bugs happen. That's a fact of life. Arguing that "we had to break
something because we were fixing a bug" is completely insane. We fix
tens of bugs every single day, thinking that "fixing a bug" means that
we can break something is simply NOT TRUE.
So bugs simply aren't even relevant to the discussion. They happen,
they get found, they get fixed, and it has nothing to do with "we
break users".
Because the only thing that matters IS THE USER.
How hard is that to understand?
Anybody who uses "but it was buggy" as an argument is entirely missing
the point. As far as the USER was concerned, it wasn't buggy - it
worked for him/her.
Maybe it worked *because* the user had taken the bug into account,
maybe it worked because the user didn't notice - again, it doesn't
matter. It worked for the user.
Breaking a user workflow for a "bug" is absolutely the WORST reason
for breakage you can imagine.
It's basically saying "I took something that worked, and I broke it,
but now it's better". Do you not see how f*cking insane that statement
is?
And without users, your program is not a program, it's a pointless
piece of code that you might as well throw away.
Seriously. This is *why* the #1 rule for kernel development is "we
don't break users". Because "I fixed a bug" is absolutely NOT AN
ARGUMENT if that bug fix broke a user setup. You actually introduced a
MUCH BIGGER bug by "fixing" something that the user clearly didn't
even care about.
And dammit, we upgrade the kernel ALL THE TIME without upgrading any
other programs at all. It is absolutely required, because flag-days
and dependencies are horribly bad.
And it is also required simply because I as a kernel developer do not
upgrade random other tools that I don't even care about as I develop
the kernel, and I want any of my users to feel safe doing the same
time.
So no. Your rule is COMPLETELY wrong. If you cannot upgrade a kernel
without upgrading some other random binary, then we have a problem.
* From `2021-06-05
<https://lore.kernel.org/all/CAHk-=wiUVqHN76YUwhkjZzwTdjMMJf_zN4+u7vEJjmEGh3recw@mail.gmail.com/>`_::
THERE ARE NO VALID ARGUMENTS FOR REGRESSIONS.
Honestly, security people need to understand that "not working" is not
a success case of security. It's a failure case.
Yes, "not working" may be secure. But security in that case is *pointless*.
* From `2011-05-06 (1/3)
<https://lore.kernel.org/all/BANLkTim9YvResB+PwRp7QTK-a5VNg2PvmQ@mail.gmail.com/>`_::
Binary compatibility is more important.
And if binaries don't use the interface to parse the format (or just
parse it wrongly - see the fairly recent example of adding uuid's to
/proc/self/mountinfo), then it's a regression.
And regressions get reverted, unless there are security issues or
similar that makes us go "Oh Gods, we really have to break things".
I don't understand why this simple logic is so hard for some kernel
developers to understand. Reality matters. Your personal wishes matter
NOT AT ALL.
If you made an interface that can be used without parsing the
interface description, then we're stuck with the interface. Theory
simply doesn't matter.
You could help fix the tools, and try to avoid the compatibility
issues that way. There aren't that many of them.
From `2011-05-06 (2/3)
<https://lore.kernel.org/all/BANLkTi=KVXjKR82sqsz4gwjr+E0vtqCmvA@mail.gmail.com/>`_::
it's clearly NOT an internal tracepoint. By definition. It's being
used by powertop.
From `2011-05-06 (3/3)
<https://lore.kernel.org/all/BANLkTinazaXRdGovYL7rRVp+j6HbJ7pzhg@mail.gmail.com/>`_::
We have programs that use that ABI and thus it's a regression if they break.
* From `2012-07-06 <https://lore.kernel.org/all/CA+55aFwnLJ+0sjx92EGREGTWOx84wwKaraSzpTNJwPVV8edw8g@mail.gmail.com/>`_::
> Now this got me wondering if Debian _unstable_ actually qualifies as a
> standard distro userspace.
Oh, if the kernel breaks some standard user space, that counts. Tons
of people run Debian unstable
* From `2019-09-15
<https://lore.kernel.org/lkml/CAHk-=wiP4K8DRJWsCo=20hn_6054xBamGKF2kPgUzpB5aMaofA@mail.gmail.com/>`_::
One _particularly_ last-minute revert is the top-most commit (ignoring
the version change itself) done just before the release, and while
it's very annoying, it's perhaps also instructive.
What's instructive about it is that I reverted a commit that wasn't
actually buggy. In fact, it was doing exactly what it set out to do,
and did it very well. In fact it did it _so_ well that the much
improved IO patterns it caused then ended up revealing a user-visible
regression due to a real bug in a completely unrelated area.
The actual details of that regression are not the reason I point that
revert out as instructive, though. It's more that it's an instructive
example of what counts as a regression, and what the whole "no
regressions" kernel rule means. The reverted commit didn't change any
API's, and it didn't introduce any new bugs. But it ended up exposing
another problem, and as such caused a kernel upgrade to fail for a
user. So it got reverted.
The point here being that we revert based on user-reported _behavior_,
not based on some "it changes the ABI" or "it caused a bug" concept.
The problem was really pre-existing, and it just didn't happen to
trigger before. The better IO patterns introduced by the change just
happened to expose an old bug, and people had grown to depend on the
previously benign behavior of that old issue.
And never fear, we'll re-introduce the fix that improved on the IO
patterns once we've decided just how to handle the fact that we had a
bad interaction with an interface that people had then just happened
to rely on incidental behavior for before. It's just that we'll have
to hash through how to do that (there are no less than three different
patches by three different developers being discussed, and there might
be more coming...). In the meantime, I reverted the thing that exposed
the problem to users for this release, even if I hope it will be
re-introduced (perhaps even backported as a stable patch) once we have
consensus about the issue it exposed.
Take-away from the whole thing: it's not about whether you change the
kernel-userspace ABI, or fix a bug, or about whether the old code
"should never have worked in the first place". It's about whether
something breaks existing users' workflow.
Anyway, that was my little aside on the whole regression thing. Since
it's that "first rule of kernel programming", I felt it is perhaps
worth just bringing it up every once in a while
..
end-of-content
..
This text is available under GPL-2.0+ or CC-BY-4.0, as stated at the top
of the file. If you want to distribute this text under CC-BY-4.0 only,
please use "The Linux kernel developers" for author attribution and link
this as source:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/plain/Documentation/process/handling-regressions.rst
..
Note: Only the content of this RST file as found in the Linux kernel sources
is available under CC-BY-4.0, as versions of this text that were processed
(for example by the kernel's build system) might contain content taken from
files which use a more restrictive license.