Re: write barriers - Was: Re: [RFC][PATCH] problem of cont_prepare_write()

[Chris Friesen <cfriesen@nortelnetworks.com>]

Anton Altaparmakov wrote:

> If I am setting two variables in sequence and it is essential that if a
> different cpu reads those variables it seems them updated in the same
> order as they were written in the C code do I need a write barrier in
> between the two?  For example:
> 
> ntfs_inode->allocated_size = 10;
> ntfs_inode->initilized_size = 10;

I believe so.  You may also need to cast them as volatile to prevent the 
compiler from reordering--can someone with more gcc knowledge than I 
state definitively whether or not it is smart enough to not reorder 
barriers?

> Should another CPU see initialized_size = 10 but allocated_size < 10 the
> ntfs driver will blow up in some places.  So does that mean I need a
> write barrier, between the two?

As above.

You may also need a read barrier to ensure that they are not 
speculatively loaded in the wrong order--could someone more knowledgable 
than I comment on that?

> If yes, do I still need it if I wrap the two settings (and all accesses)
> with a spin lock?  And in particular with a rw-spinlock?  For example:
> 
> write_lock_irqsave(&ntfs_inode->size_lock, flags);
> ntfs_inode->allocated_size = 10;
> ntfs_inode->initilized_size = 10;
> write_unlock_irqrestore(&ntfs_inode->size_lock, flags);
> 
> Do I still need a write barrier or does the spinlock imply it already?

I believe the spinlock implies it.

> Thanks a lot in advance and apologies for the stupid(?) questions...

Not stupid.  Concurrency is hard.

Chris