[NCLUG] serialization question

[Matthew Wilcox]

On Wed, Jun 26, 2002 at 02:42:36PM -0600, Larry Butler wrote:
> First, I'm looking at some code that suggests that calls to open and release 
> are serialized in a given driver.  Please confirm or deny.

This is a struct file_operations, right?  If so, the locking rules are
(tersely) documented in Documentation/filesystems/Locking.  I believe that
open and release are _not_ serialised wrt each other and you will need to
perform your own form of exclusion.

> Second, if I deny multiple openers and don't sleep in write or ioctl, can I 
> expect that calls to write and ioctl to be serialized?  I would expect not 
> (even with the restriction on multiple openers) since a process can fork and 
> its child could write while the parent is writing.  However, I have at least 
> one example of code that does some pretty involved poking at hardware without 
> any precautions.

Definitely unsafe.  fork() is one case where an fd can be shared,
and another is unix domain socket fd passing.  You're going to need
a spin_lock_irq()/spin_unlock_irq() style construct while you poke at
device registers.

> Third, where in the Rubini book are the answers?  I must be blind.  I can't 
> find them.

Hmm.. not sure.  I've only looked at the bits of the 2nd edition Rubini
book where I want to make changes (to see what driver authors are being
told to rely on ;-)

Best of luck.

-- 
Revolutions do not require corporate support.