Le 23/01/2020 à 19:02, Linus Torvalds a écrit :

On Thu, Jan 23, 2020 at 4:59 AM Christophe Leroy
[off-list ref] wrote:

quoted

On 32 bits powerPC (book3s/32), only write accesses to user are
protected and there is no point spending time on unlocking for reads.

Honestly, I'm starting to think that 32-bit ppc just needs to look
more like everybody else, than make these changes.

Well, beside ppc32, I was also seen it as an opportunity for the modern 
ppc64. On it, you can unlock either read or write or both. And this is 
what is done for get_user() / put_user() and friends: unlock only reads 
for get_user() and only writes for put_user().

Could also be a compromise between performance and security: keeping 
reads allowed at all time and only protect against writes on modern 
architectures which support it like ppc64.

We used to have a read/write argument to the old "verify_area()" and
"access_ok()" model, and it was a mistake. It was due to odd i386 user
access issues. We got rid of it. I'm not convinced this is any better
- it looks very similar and for odd ppc access issues.

I'm going to leave it aside, at least for the time being, and do it as a 
second step later after evaluating the real performance impact. I'll 
respin tomorrow in that way.

But if we really do want to do this, then:

Indeed I took the idea from a discussion in last Octobre (Subject: 
"book3s/32 KUAP (was Re: [PATCH] Convert filldir[64]() from __put_user() 
to unsafe_put_user())" )

https://lore.kernel.org/lkml/87h84avffi.fsf@mpe.ellerman.id.au/ (local)

quoted

Add an argument to user_access_begin() to tell when it's for write and
return an opaque key that will be used by user_access_end() to know
what was done by user_access_begin().

You should make it more opaque than "unsigned long".

Also, it shouldn't be a "is this a write". What if it's a read _and_ a
write? Only a write? Only a read?

Indeed that was more: does it includes a write. It's either RO or RW

Christophe