Re: [PATCH net] tls: Skip tls_append_frag on zero copy size

[Jakub Kicinski <kuba@kernel.org>]

On Thu, 21 Apr 2022 12:47:18 +0300 Maxim Mikityanskiy wrote:
> On 2022-04-18 17:56, Maxim Mikityanskiy wrote:
> > On 2022-04-14 13:28, Jakub Kicinski wrote:  
> >> I appreciate you're likely trying to keep the fix minimal but Greg
> >> always says "fix it right, worry about backports later".
> >>
> >> I think we should skip more, we can reorder the mins and if
> >> min(size, rec space) == 0 then we can skip the allocation as well.  
> > 
> > Sorry, I didn't get the idea. Could you elaborate?
> > 
> > Reordering the mins:
> > 
> > copy = min_t(size_t, size, max_open_record_len - record->len);
> > copy = min_t(size_t, copy, pfrag->size - pfrag->offset);
> > 
> > I assume by skipping the allocation you mean skipping 
> > tls_do_allocation(), right? Do you suggest to skip it if the result of 
> > the first min_t() is 0?
> > 
> > record->len used in the first min_t() comes from ctx->open_record, which 
> > either exists or is allocated by tls_do_allocation(). If we move the 
> > copy == 0 check above the tls_do_allocation() call, first we'll have to 
> > check whether ctx->open_record is NULL, which is currently checked by 
> > tls_do_allocation() itself.
> > 
> > If open_record is not NULL, there isn't much to skip in 
> > tls_do_allocation on copy == 0, the main part is already skipped, 
> > regardless of the value of copy. If open_record is NULL, we can't skip 
> > tls_do_allocation, and copy won't be 0 afterwards.
> > 
> > To compare, before (pseudocode):
> > 
> > tls_do_allocation {
> >      if (!ctx->open_record)
> >          ALLOCATE RECORD
> >          Now ctx->open_record is not NULL
> >      if (!sk_page_frag_refill(sk, pfrag))
> >          return -ENOMEM
> > }
> > handle errors from tls_do_allocation
> > copy = min(size, pfrag->size - pfrag->offset)
> > copy = min(copy, max_open_record_len - ctx->open_record->len)
> > if (copy)
> >      copy data and append frag
> > 
> > After:
> > 
> > if (ctx->open_record) {
> >      copy = min(size, max_open_record_len - ctx->open_record->len)
> >      if (copy) {
> >          // You want to put this part of tls_do_allocation under if (copy)?
> >          if (!sk_page_frag_refill(sk, pfrag))
> >              handle errors
> >          copy = min(copy, pfrag->size - pfrag->offset)
> >          if (copy)
> >              copy data and append frag
> >      }
> > } else {
> >      ALLOCATE RECORD
> >      if (!sk_page_frag_refill(sk, pfrag))
> >          handle errors
> >      // Have to do this after the allocation anyway.
> >      copy = min(size, max_open_record_len - ctx->open_record->len)
> >      copy = min(copy, pfrag->size - pfrag->offset)
> >      if (copy)
> >          copy data and append frag
> > }
> > 
> > Either I totally don't get what you suggested, or it doesn't make sense 
> > to me, because we have +1 branch in the common path when a record is 
> > open and copy is not 0, no changes when there is no record, and more 
> > repeating code hard to compress.
> > 
> > If I missed your idea, please explain in more details.  
> 
> Jakub, is your comment still relevant after my response? If not, can the 
> patch be merged?

I'd prefer if you refactored the code so tls_push_data() looks more
natural. But the patch is correct so if you don't want to you can
repost.

Sorry for the delay.