Re: [PATCH v4 3/4] iov_iter: add copy_page_to_iter_atomic()

[Baoquan He <bhe@redhat.com>]

On 03/22/23 at 11:06am, Lorenzo Stoakes wrote:
> On Wed, Mar 22, 2023 at 10:32:47AM +0000, Lorenzo Stoakes wrote:
> > > I am a little confused about the name of this new function. In its
> > > conterpart, copy_page_from_iter_atomic(), kmap_atomic()/kunmpa_atomic()
> > > are used. With them, if CONFIG_HIGHMEM=n, it's like below:
> >
> > The reason for this is that:-
> >
> > 1. kmap_atomic() explicitly states that it is now deprecated and must no longer
> >    be used, and kmap_local_page() should be used instead:-
> >
> >  * kmap_atomic - Atomically map a page for temporary usage - Deprecated!
> >
> >  * Do not use in new code. Use kmap_local_page() instead.
> >
> > 2. kmap_local_page() explicitly states that it can be used in any context:-
> >
> >  * Can be invoked from any context, including interrupts.
> >
> > I wanted follow this advice as strictly as I could, hence the change. However,
> > we do need preemption/pagefaults explicitly disabled in this context (we are
> > happy to fail if the faulted in pages are unmapped in meantime), and I didn't
> > check the internals to make sure.
> >
> > So I think for safety it is better to use k[un]map_atomic() here, I'll respin
> > and put that back in, good catch!
> >
> 
> Actually, given we have preemption disabled due to the held spinlock, I think
> it'd be better to add a copy_page_to_iter_nofault() that uses
> copy_to_user_nofault() which will disable pagefaults thus have exactly the
> equivalent behaviour, more explicitly and without the use of a deprecated
> function.

Sounds a great idea, that let us be able to avoid using kmap_atomic.

> 
> > >
> > > static inline void *kmap_atomic(struct page *page)
> > > {
> > >         if (IS_ENABLED(CONFIG_PREEMPT_RT))
> > >                 migrate_disable();
> > >         else
> > >                 preempt_disable();
> > >         pagefault_disable();
> > >         return page_address(page);
> > > }
> > >
> > > But kmap_local_page() is only having page_address(), the code block
> > > between kmap_local_page() and kunmap_local() is also atomic, it's a
> > > little messy in my mind.
> > >
> > > static inline void *kmap_local_page(struct page *page)
> > > {
> > >         return page_address(page);
> > > }
> > >
> > > > +	char *p = kaddr + offset;
> > > > +	size_t copied = 0;
> > > > +
> > > > +	if (!page_copy_sane(page, offset, bytes) ||
> > > > +	    WARN_ON_ONCE(i->data_source))
> > > > +		goto out;
> > > > +
> > > > +	if (unlikely(iov_iter_is_pipe(i))) {
> > > > +		copied = copy_page_to_iter_pipe(page, offset, bytes, i);
> > > > +		goto out;
> > > > +	}
> > > > +
> > > > +	iterate_and_advance(i, bytes, base, len, off,
> > > > +		copyout(base, p + off, len),
> > > > +		memcpy(base, p + off, len)
> > > > +	)
> > > > +	copied = bytes;
> > > > +
> > > > +out:
> > > > +	kunmap_local(kaddr);
> > > > +	return copied;
> > > > +}
> > > > +EXPORT_SYMBOL(copy_page_to_iter_atomic);
> > > > +
> > > >  static void pipe_advance(struct iov_iter *i, size_t size)
> > > >  {
> > > >  	struct pipe_inode_info *pipe = i->pipe;
> > > > --
> > > > 2.39.2
> > > >
> > >
>