Re: [Qemu-devel] [PATCH] linux-user EFAULT implementation

[Fabrice Bellard <fabrice@bellard.org>]

Paul Brook wrote:
(...]
> Using g2h directly is bad. g2h is an implementation detail of one particular 
> memory model.
> 
> The whole point of the lock_user abstraction (or a similar copy_from_user 
> abstraction) is that almost none of the code cares how "user" memory is 
> accessed. One of the long-term goals of this abstraction is to allow the 
> softmmu code to be used with userspace emulation. In this case a region may 
> be split across multiple discontiguous host pages.

I confirm that I expected g2h() to be completely removed in case the 
Linux user access API is used.

> The reason I used a locking paradigm rather than a copying one is that it 
> allows a zero-copy implementation in the common case. I've no strong 
> objections to a copying interface, however it must be implementation 
> agnostic.

Since you agree for the copying interface, I suggest to convert all the 
code to it. The exact implementation of access_ok(), copy_to_user(), 
copy_from_user()... will be easily modifiable if one day softmmu user 
mode is needed.

In the only specific case of potentially big unbounded memory areas 
(such as read/write buffers), a different API must be used where it is 
possible to get a pointer to user pages (e.g. ptr = lock_user_page(addr, 
rw_mode); if (!ptr) return -EFAULT; unlock_user_page(ptr)), so that no 
copy is performed while being able to handle any user page remapping.

Regards,

Fabrice.