Re: [PATCH v3 41/55] iscsi: Assume "sendpage" is okay in iscsi_tcp_segment_map()

[David Howells]

Fabio M. De Francesco <fmdefrancesco@xxxxxxxxx> wrote:

> > -	if (recv) {
> > -		segment->atomic_mapped = true;
> > -		segment->sg_mapped = kmap_atomic(sg_page(sg));
> > -	} else {
> > -		segment->atomic_mapped = false;
> > -		/* the xmit path can sleep with the page mapped so use 
> kmap */
> > -		segment->sg_mapped = kmap(sg_page(sg));
> > -	}
> > -
> > +	segment->atomic_mapped = true;
> > +	segment->sg_mapped = kmap_atomic(sg_page(sg));
> 
> As you probably know, kmap_atomic() is deprecated.
> 
> I must admit that I'm not an expert of this code, however, it looks like the 
> mapping has no need to rely on the side effects of kmap_atomic() (i.e., 
> pagefault_disable() and preempt_disable() - but I'm not entirely sure about 
> the possibility that preemption should be explicitly disabled along with the 
> replacement with kmap_local_page()). 
> 
> Last year I've been working on several conversions from kmap{,_atomic}() to 
> kmap_local_page(), however I'm still not sure to understand what's happening 
> here...
> 
> Am I missing any important details? Can you please explain why we still need 
> that kmap_atomic() instead of kmap_local_page()? 

Actually, it might be worth dropping segment->sg_mapped and segment->data and
only doing the kmap_local when necessary.

And this:

			struct msghdr msg = { .msg_flags = flags };
			struct kvec iov = {
				.iov_base = segment->data + offset,
				.iov_len = copy
			};

			r = kernel_sendmsg(sk, &msg, &iov, 1, copy);

should really be using struct bvec, not struct kvec - then the mapping isn't
necessary.  It looks like this might be the only place the mapping is used,
but I'm not 100% certain.

David