Re: [PATCH 0/7] isdn4linux: add drivers for Siemens Gigaset ISDN DECT PABX

[Tilman Schmidt <tilman@imap.cc>]

[-- Attachment #1: Type: text/plain, Size: 3425 bytes --]

On Fri, 03 Mar 2006 07:53:28 +0100, Arjan van de Ven wrote:

> On Fri, 2006-03-03 at 00:03 +0100, Tilman Schmidt wrote:

>> So you are saying that, for example
>>
>> 	spin_lock_irqsave(&cs->ev_lock, flags);
>> 	head = cs->ev_head;
>> 	tail = cs->ev_tail;
>> 	spin_unlock_irqrestore(&cs->ev_lock, flags);
>>
>> is (mutatis mutandis) actually cheaper than
>>
>> 	head = atomic_read(&cs->ev_head);
>> 	tail = atomic_read(&cs->ev_tail);
>
> atomic_read is special since it's not actually an atomic operation ;)
> but.. think about it: you do 2 atomic reads, however there is ZERO
> guarantee that the reads are atomic with respect to eachother; eg your
> head and tail are not an atomic "snapshot" of these 2 variables!

That's not a problem. It's a ringbuffer. It doesn't need an atomic
snapshot of the reading and writing pointers together. Nothing breaks
if a reader advances the read pointer while a writer is holding a
local copy of it, or vice versa. The only thing we have to guard
against is the result of an individual read operation being corrupted
by a parallel write.

So what's better in that case? Should we change these from atomic to
spinlocked or not?

[#define IFNULL*]
>> Ok, these were mainly debugging aids. We'll just drop them and let the
>> oops mechanism catch the (hopefully non-existent) remaining cases of
>> pointers being unexpectedly NULL.
>
> you can also use WARN_ON() and BUG_ON() for that, you then get a more
> readable oops message (with filename and line information)

Actually, we won't. The IFNULL* macros were not only printing a message,
but also taking evasive action in order to avoid dereferencing the NULL
pointer. To achieve the same with WARN_ON() would require four lines of
code for each occurrence, which IMHO is too much code clutter for a class
of bugs which should be largely eradicated by now anyway.

>> >> +void gigaset_dbg_buffer(enum debuglevel level, const unsigned char *msg,
>> >> +			size_t len, const unsigned char *buf, int from_user)
>> >
>> > such "from_user" parameter is highly evil, and also breaks sparse and
>> > friends.. (btw please run sparse on the code and fix all warnings)
>>
>> Are you referring to anything in particular? We do run sparse regularly,
>> and it did not emit any warnings for the submitted version, not even for
>> this function. (But heaps of them for other parts of the kernel, if you
>> pardon the remark.)
>
> msg should have the __user atribute here since it can be in userspace...
> sometimes. It is the "sometimes" that is the bad idea!

That's understood and will be fixed. I was just wondering whether your
remark in parentheses was prompted by any particular sparse warnings you
wanted us to fix and which for some reason we hadn't seen?

> (GFP_ATOMIC is like borrowing from the VM, the VM will be in slight
> imbalance afterwards. With GFP_KERNEL you allow the kernel to fix this
> imbalance. A slight imbalance is fine and not a problem. Especially if
> you give it the memory back soon. But if the imbalance can accumulate,
> for example because you keep allocating and free the memory much later,
> it can become a problem)

Thanks muchly for that very lucid explanation. I see much clearer now
in that area! :-)

Regards
Tilman

--
Tilman Schmidt                    E-Mail: tilman@imap.cc
Bonn, Germany
Diese Nachricht besteht zu 100% aus wiederverwerteten Bits.
Ungeoeffnet mindestens haltbar bis: (siehe Rueckseite)

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 250 bytes --]