Why kernelspace sucks?

Why kernel space sucks
(Or: Deconstructing two myths)
(Or: An abridged history of kernel-userspace interface blunders...)

Michael Kerrisk
OpenFest 2011
© 2011
5 Nov 2011
mtk@man7.org
Sofia, Bulgaria
http://man7.org/

Who am I?
● Professionally: programmer (primarily); also teacher
and writer
● Working with UNIX and Linux for nearly 25 years
● Linux man-pages maintainer since 2004
● 117 releases to date
● written or cowritten ~250 of ~950 man pages
● lots of API testing, many bug reports
● Author of a book on the kernel-userspace API
● IOW: I've spent a lot of time looking at the interface

2
man7.org

Intro: Why Userspace Sucks
● Paper/talk by Dave Jones of Red Hat
● First presented at Ottawa LS 2006
● A lead-in to deconstructing a couple of myths
● Why Userspace Sucks → WUSS

● http://www.kernel.org/doc/ols/2006/ols2006v1-pages-441-450.pdf
● http://www.codemonkey.org.uk/projects/talks/ols2k6.tar.gz
● http://lwn.net/Articles/192214/

3
man7.org

Motivation for WUSS
● We (kernel developers) have created a
magnificently performant kernel
● But, can we make it better?
● Why does it take so long to boot, start applications,
and shut down?
● Why does my idle laptop consume so much battery
power?

4
man7.org

Findings from WUSS
● DJ starts instrumenting the kernel, and finds...
● Boot up: 80k stat(), 27k open(), 1.4k exec()
● Shutdown: 23k stat(), 9k open()
● Userspace programmers wreck performance
doing crazy things!
● open() and reparse same file multiple times!
● read config files for many devices not even present!
● stat() (or even open()) 100s of files they never need
● timers triggering regular unnecessary wake-ups

5
man7.org

Conclusions from WUSS
● Room for a lot of improvement in userspace!
● Userspace programmers should be aware of
and using trace and analysis tools
● (perf, LTTng, ftrace, systemtap, strace, valgrind,
powertop, etc.)

6
man7.org

Kernelspace Userspace

“We (kernel developers) are much smarter than
those crazy userspace programmers”

Kernelspace Userspace

Something's wrong with this
picture...

Let's question a couple of myths...
● Myth 1: Kernel programmers (can) always get
things right (in the end)
● Myth 2: Code is always the best way to
contribute to Free Software

9
man7.org

Myth 1
Kernel programmers
(can) always get things right
(in the end)

Except, there's (at least) one place
where they don't: the interface

The kernel-userspace interface
● Application programming interface (API)
presented by kernel to userspace programs
● System calls (← I'll focus here)
● Pseudo-file systems (/proc, /sys, etc.)
● ioctl() interfaces (device drivers)
● Netlink sockets
● Obscure pieces (AUXV, VDSO, ...)

11
man7.org

API designs must be done
right first time

Why must APIs be right first time?
● Code fixes != API fixes

13
man7.org

Why is fixing APIs so hard?
● Usually, “fixing” an interface means breaking
the interface for some applications

“We care about user-space interfaces to an insane
degree. We go to extreme lengths to maintain even
badly designed or unintentional interfaces.
Breaking user programs simply isn't acceptable.”
[LKML, Dec 2005]

14
man7.org

We have to live with our mistakes!

So: any API mistake by kernel
hackers creates pain that thousands
of userspace programmers must live
with for decades

So, what does it mean
to get an API right?

Doing (kernel-userspace) APIs right
● Properly designed and implemented API should:
● be bug free!
● have a well thought out design
– simple as possible (but no simpler)
– easy to use / difficult to misuse
● be consistent with related/similar APIs
● integrate well with existing APIs
– e.g., interactions with fork(), exec(), threads, signals, FDs?
● be as general as possible
● be extensible, where needed; accommodate future growth trends
● adhere to relevant standards (as far as possible)
● be as good as, or better than, earlier APIs with similar functionality
● be maintainable over time (a multilayered question)

17
man7.org

So how do kernel
developers score?
(BSDers, please laugh quietly)

Bugs
● utimensat(2) [2.6.22]
● Set file timestamps
● Multiple bugs!
– http://linux-man-pages.blogspot.com/2008/06/whats-wrong-with-kernel-userland_30.html

● Fixed in 2.6.26
● signalfd() [2.6.22]
● Receive signals via a file descriptor
● Didn't correctly obtain data sent with sigqueue(2)
● Fixed in 2.6.25

20
man7.org

Bugs
● Examples of other interfaces with significant,
easy to find bugs at release:
● inotify [2.6.13]
● splice() [2.6.17] (http://marc.info/?l=linux-mm&m=114238448331607&w=2)
● timerfd [2.6.22] (http://marc.info/?l=linux-kernel&m=118517213626087&w=2)

21
man7.org

Bugs—what's going on?
● There's a quality control issue; way too many
bugs in released interfaces
● Pre-release testing insufficient and haphazard:
● Too few testers (maybe just kernel developer)
● No unit tests
● Insufficient test coverage
● No clear specification against which to test
● Even if bug is fixed, users may still need to care
● special casing for kernel versions

22
man7.org

Code it now, think about it later
● Vanishing arguments:
● readdir(2) ignores count argument
● getcpu(2) [2.6.19] ignores tcache argument
● epoll_create() [2.6] ignores size arg. (must be > 0)
since 2.6.8
● Probably, argument wasn't needed to start with
● Or: recognized as a bad idea and made a no-op

24
man7.org

● futimesat() [2.6.16]
● Extends utimes()
● Proposed for POSIX.1-2008
● Implemented on Linux
● POSIX.1 members realize API is insufficient
→ standardized different API
● utimensat() added in Linux 2.6.22

25
man7.org

● Dec 2003: Linux 2.6 added epoll_wait()
● File descriptor monitoring
– (improves on select())
● Nov 2006, 2.6.19 added epoll_pwait() to allow
manipulation of signal mask during call
● But, already in 2001, POSIX specified pselect() to
fix analogous, well-known problem in select()

26
man7.org

Interface inconsistencies
● mlock(start, length):
● Round start down to page size
● Round length up to next page boundary
● mlock(4000, 6000) affects bytes 0..12287
● remap_file_pages(start, length, ...) [2.6]:
● Round start down to page boundary
● Round length down to page boundary(!)
● remap_file_pages(4000, 6000, ...) ? → 0..4095
● User expectation: similar APIs should behave
similarly

28
man7.org

Confusing the users
● Various system calls allow one process to
change attributes of another process
● e.g., setpriority(), ioprio_set(), migrate_pages(),
prlimit()
● Unprivileged calls require credential matches:
● Some combination of caller's UIDs/GIDs matches
some combination of target's UIDs/GIDs

29
man7.org

Confusing the users
● But, much inconsistency; e.g.:
● setpriority(): euid == t-uid || euid == t-euid
● ioprio_set(): uid == t-uid || euid == t-uid
● migrate_pages(): uid == t-uid || uid == t-suid || euid == t-uid ||
euid == t-suid
● prlimit(): (uid == t-uid && uid == t-euid && uid == t-suid) &&
(gid == t-gid && gid == t-guid && gid == t-sgid) !!
● Inconsistency may confuse users into writing
bugs
● Potentially, security related bugs!
● http://linux-man-pages.blogspot.com/2010/11/system-call-credential-checking-tale-of.html

30
man7.org

Is the interface sufficiently general?
● 2.6.22 added timerfd(ufd, flags, utimerspec)
● Create timer that notifies via a file descriptor
● But API didn't allow user to:
● Retrieve previous value when setting new timer value
● Do a “get” to retrieve time until next expiration
– http://marc.info/?l=linux-kernel&m=118517213626087&w=2
– http://lwn.net/Articles/245533/
● Older APIs ([gs]etitimer(), POSIX timers) did
provide this functionality!

32
man7.org

Is the interface sufficiently general?
● Solution:
● timerfd() disabled in kernel 2.6.23
● 2.6.25 did it right:
– timerfd_create(), timerfd_settime(), timerfd_gettime()
– (API analogous to POSIX timers)
● Was an ABI breakage, but
● Only in a single kernel version
● Original API was never exposed via glibc

33
man7.org

Are we learning
from the past?

Are we learning from past mistakes?
● Dnotify [2.4]
● Directory change notification API
● Many problems
● So, we added inotify [2.6.13]
● But, inotify still doesn't get it all right
● Now [2.6.37] we have yet another API, fanotify
● Designed for virus scanners
● Adds some functionality
● Doesn't provide all functionality of inotify
● Couldn't we have had a new API that did everything?

35
man7.org

Is the interface extensible?
● Too often, an early API didn't allow for
extensions
● Common solution is a new API, with a flags arg:
● umount() → umount2() [2.2]
● epoll_create() [2.6] → epoll_create2() [2.6.27]
● futimesat() [2.6.16] → epoll_create2() [2.6.22]
● signalfd() [2.6.22] → signalfd4() [2.6.27]
● When adding a new API, consider adding an
(unused) flags argument to allow extensions

37
man7.org

Futureproofing
● Suppose a syscall has a flags bit-mask arg.
● Implementation should always have check like:
if (flags & ~(FL_X | FL_Y))
return -EINVAL;
// Only allow caller to specify flags
// bits that have a defined meaning

● Without this check, interface is “loose”

38
man7.org

Futureproofing
● Suppose user makes a call of form:
syscallxyz(-1); // flags has all bits set

● If implementer later adds FL_Z, an ABI
breakage occurs for user's code
● Conversely: user has no way of checking if a
kernel implements FL_Z
● Many system calls lack this kind of check!
● Linux 3.2 examples: sigaction(sa.sa_flags), recv(),
send(), clock_nanosleep(), msgrcv(), msgget(),
semget(), shmget(), shmat(), semop(sops.sem_flg)
39
man7.org

Futureproofing
● Should checks be added after the fact?
● e.g., umount2() [2.2] added check in 2.6.34;
timerfd_settime() [2.6.25] added check in 2.6.29
● But adding check can also create ABI breakage
● Apps get errors where previously they did not
● Loose APIs allow the user to define interface
● Worst case: can't add new flags values to interface

40
man7.org

Futureproofing failures
● 16 bits is enough for UIDs/GIDs...
● 2.4: 32-bit UIDs/GIDs
● 32 bits is enough for file offsets
● Okay, it was 1991, but Moore's law...
● 2.4: 64-bit file offsets
● So we have
● oldstat(), stat(), stat64()
● chown(), chown32()
● open(), open64()
● and so on
41
man7.org

When good ideas go astray
● Traditional UNIX gives root all privileges
● All or nothing is risky!
● Linux divides root privileges into separate pieces:
● CAP_AUDIT_CONTROL, CAP_AUDIT_WRITE, CAP_CHOWN, CAP_DAC_OVERRIDE, CAP_DAC_READ_SEARCH,
CAP_FOWNER, CAP_FSETID, CAP_IPC_LOCK, CAP_IPC_OWNER, CAP_KILL, CAP_LEASE,
CAP_LINUX_IMMUTABLE, CAP_MAC_ADMIN, CAP_MAC_OVERRIDE, CAP_MKNOD, CAP_NET_ADMIN,
CAP_NET_BIND_SERVICE, CAP_NET_BROADCAST, CAP_NET_RAW, CAP_SETFCAP, CAP_SETGID, CAP_SETPCAP,
CAP_SETUID, CAP_SYSLOG, CAP_SYS_ADMIN, CAP_SYS_BOOT, CAP_SYS_CHROOT, CAP_SYS_MODULE,
CAP_SYS_NICE, CAP_SYS_PACCT, CAP_SYS_PTRACE, CAP_SYS_RAWIO, CAP_SYS_RESOURCE, CAP_SYS_TIME,
CAP_SYS_TTY_CONFIG, CAP_WAKE_ALARM

● Great! But which capability do I use for my new
feature?
● Hmmm, CAP_SYS_ADMIN looks good
● CAP_SYS_ADMIN, the new root, 231 uses in 3.2
43
man7.org

Needlessly breaking portability
● sched_setscheduler()
● POSIX says: successful call must return previous
policy
● Linux: successful call returns 0
● No good reason for this inconsistency
● Developers must special case code for Linux

45
man7.org

Actually,
it wasn't just us...

We're just traditionalists...
● These kinds of problems predate Linux:
● API of System V IPC is awful!
● Semantics of fcntl() locks when FD is closed render
locks useless for libraries
● select() modifies FD sets in place, forcing
reinitialization inside loops
– poll() gets it right: uses separate input and output args
● and so on...

47
man7.org

Why do these API problems
keep happening?
● Excessive focus on code as primary
contribution of value for a software project
● Poor feedback loop between developers and
users

Myth 2
Code is always the best way
to contribute to Free Software

“Show me the code!”

But anyone can write code,
and if the design is good
but the code is bad,
the code can usually be fixed

“Show me the code!”

Sometimes,
other sentences are more appropriate,
and encourage contributions that are
as (or more) valuable

“Show me the users'
requirements!”

“Show me the users' requirements”
● Does the API serve needs of multiple users, or
is it just one developer scratching an itch?
● Beware of limited perspectives!
● Is the API designed for generality?
● Is the API extensible for possible future
requirements?

55
man7.org

“Show me the design
specification / documentation!”

“Show me the design spec. / documentation!”
● How do we know if implementation deviates
from intention?
● What shall we code our tests against?
● Writing documentation turns out often to be a
natural sanity check for design
● A decent man page suffices
● Many of the bugs mentioned earlier were found
while writing man pages...
● It's all a question of when it's written...

57
man7.org

“Show me the design review!”

“Show me the design review!”
● Did other people actually review your design?
● Is the API:
● as simple as possible?
● easy to use / difficult to misuse?
● consistent with related/similar APIs?
● well integrated with existing APIs?
● as general as possible
● extensible?
● following standards, where relevant?
● at least as good as earlier APIs with similar functionality?
● maintainable?
59
man7.org

“Show me the tests!”
● Did you (the developer) write some tests?
● More importantly: did someone else write some
tests?
● Do the tests cover all reasonable cases?
● Do you test for unreasonable cases?
● Do unexpected inputs generate suitable error returns?
● While writing tests, did you find the interface easy
to use / difficult to misuse? (Did you consequently
make some design changes?)
● What bugs did you discover during testing?
61
man7.org

Finally...
● If you're a potential contributor, don't fall into the
trap of believing that code is the only (or best)
vehicle for contribution
● As a maintainer, are you letting your project
down by failing to encourage these other types
of contribution?

Thanks!
(slides up soon at http://man7.org/conf/)

Michael Kerrisk Linux man-pages project
mtk@man7.org mtk.manpages@gmail.com
http://man7.org/ http://man7.org/linux/man-pages/

Mamaku (Black Tree Fern) image (c) Rob Suisted

New Zealand Blue Penguin (Kororā)
Image (c) livescience.com
naturespic.com

No Starch Press, 2010
http://man7.org/tlpi/

Why kernelspace sucks?

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Why kernelspace sucks? (20)

More from OpenFest team (20)

Recently uploaded (20)

Why kernelspace sucks?