Index Home About Blog
Newsgroups: fa.linux.kernel
From: Linus Torvalds <>
Subject: Re: EFI partition code broken..
Original-Message-ID: <>
Date: Sun, 7 Nov 2004 22:42:38 GMT
Message-ID: <>

On Sun, 7 Nov 2004, Matt Domsch wrote:
> Another train of thought, and copying gregkh for inspiration.  Is there
> any way to know which devices lie about their size, and fix that with
> quirk code in the device discovery routines?

The USB layer actually has some quirks like this, but I think that's a
bug waiting to happen.

The thing is, if you start doing quirks, you _are_ screwed in the end.
Don't do it. It's not just a maintenance nightmare, it's fundamentally
wrong. It fundamentally takes the approach of "you have to have a kernel
that is two years newer than the hardware you have", which is an approach
that I just find incredibly broken.

Quirks work slightly better in practice for stuff that seldom changes,
and/or where we have fairly good vendor support. So CPU's, for example,
are largely ok with quirks (aka "errata"). But random regular devices?
Please no.

Side note: the USB storage stuff has historically had tons of quirks,
largely because the SCSI layer used to do crap-all to try to be sane. The
SCSI layer historically only cared about high-end devices, and then the
USB storage model clashed pretty hard with the old SCSI layer belief that
standards are something that people follow etc.

Happily, most of those quirks are hopefully stale these days, because the
SCSI layer has been slowly converted to the idea that you don't use every
documented feature under the sun just because it exists.

So I'm trying to make for _fewer_ quirks rather than more of them.


From: Linus Torvalds <>
Newsgroups: fa.linux.kernel
Subject: Re: [GIT PATCH] USB autosuspend fixes for 2.6.23-rc6
Date: Thu, 13 Sep 2007 16:44:50 UTC
Message-ID: <fa.oCwMDFhhTXL9WgIzmuNyihbHt/>

On Thu, 13 Sep 2007, Alan Stern wrote:
> But mainly it's a question of maintenance and modification.  Kernel
> developers don't really enjoy maintaining black- or whitelists of
> devices, together with all the work involved in sorting through the
> issues when somebody posts an email saying "My device doesn't work!".


In general, I think the USB blacklist/whitelists are generally a sign of
some deeper bug.

We used to have a lot of those things due to simply incorrect SCSI
probing, causing devices to lock up because Linux probed them with bad or
unexpected modepages etc. I suspect we still have old blacklist entries
from those days that just never got cleaned up, because nobody ever dared
remove the blacklist entry.

We should strive to make the default behaviour be so safe that we never
need a black-list (or a whitelist), and basically consider blacklists to
be not a way to "fix up a device", but a way to avoid some really serious
AND *RARE* error.

The moment you have lots of devices having the same blacklist entry,
that's a sign that the blacklist is wrong, and the subsystem itself is
likely doing something bad!

So, I would seriously suggest:
 - look at USB quirks that have more than ten entries (and the entries
   aren't just the exact same device in various guises)
 - start considering that feature to be something that is known broken,
   and shouldn't be done AT ALL by default.
 - have some way to enable some extension on a device-by-device basis from
   the /sysfs interface, and then users can enable those things on their
   own with a graphical interface or something (or using whitelists in
   user space saying "ok, this device can actually do this")

It looks like the current situation now is that the latest autosuspend
patches did basically everything but the last point.

Btw, this is in no way just an AUTOSUSPEND issue. The USB layer has a
*lot* of these quirks. They are often called "UNUSUAL_DEV()", but the
thing is, some of those things seem to be so usual that the naming is
dubious, and thus calling it a "quirk" or "unusual" is pretty dubious too.

For example, why do we have that US_FL_MAX_SECTORS_64 at all? The fact
that some USB device is broken with more than 64 sectors would seem to
indicate that Windows *never* does more than 64 sectors, and that in turn
means that pretty much *no* devices have ever been tested with anything

So why not make the 64 sector limit be the default? Get rid of the quirk:
we already allow people to override it in /sys if they really want to, but
realistically, it's probably not going to make any difference what-so-ever
for *any* normal load. So we seem to have a quirk that really doesn't buy
us anything but headache.

Other quirks worth looking at (but likely unfixable) are:
	Does this really matter? Can we not just always do the
	US_FL_IGNORE_RESIDUE thing? Windows must not be doing what we're
	This is a generic SCSI issue, not a USB one, and maybe there are
	better solutions to it. Are we perhaps doing something wrong? Is
	there some patterns we haven't seen? Why do we need this, when
	presumably Windows does not?
	At least a few of these seem to indicate that the real problem
	could be detected dynamically ("device reports Sub=ff") rather
	than with a quirk. Quirks are unmaintainable (and change), but
	noticing when devices return impossible values and going into a
	"safe mode" is just defensive programming.

> Maybe you're concerned about propagating updates as painlessly as
> possible -- if the whitelist is in the kernel then every kernel release
> would include an update.  But in userspace it's possible to do updates
> even more quickly and painlessly.  For example, there could be a
> network server available for both interactive lookups and automatic
> queries from HAL.

For a lot of these things, you probably do not need a whitelist *at*all*!

IOW, just default to something safe (the 64 sector example), and then
perhaps allow people to explicitly play with their settings in a hardware
manager. People actually tend to *like* being able to tweak meaningless
things, and it makes them feel in control. So you'd have the Gentoo people
who want to optimize their iPod access times by 0.2% by raising the
maximum sector number - good for them! They'll feel empowered, and if it
stops working, they know it was because of something *they* did.

So at least in some cases, I think we should "default to stupid, but give
users rope".


Index Home About Blog