Re: [PATCH v1 1/3] net: Introduce sk_use_task_frag in struct sock.

[Eric Dumazet]

On Fri, Dec 9, 2022 at 1:09 PM Paolo Abeni <pabeni@xxxxxxxxxx> wrote:
>
> On Mon, 2022-11-21 at 08:35 -0500, Benjamin Coddington wrote:
> > From: Guillaume Nault <gnault@xxxxxxxxxx>
> >
> > Sockets that can be used while recursing into memory reclaim, like
> > those used by network block devices and file systems, mustn't use
> > current->task_frag: if the current process is already using it, then
> > the inner memory reclaim call would corrupt the task_frag structure.
> >
> > To avoid this, sk_page_frag() uses ->sk_allocation to detect sockets
> > that mustn't use current->task_frag, assuming that those used during
> > memory reclaim had their allocation constraints reflected in
> > ->sk_allocation.
> >
> > This unfortunately doesn't cover all cases: in an attempt to remove all
> > usage of GFP_NOFS and GFP_NOIO, sunrpc stopped setting these flags in
> > ->sk_allocation, and used memalloc_nofs critical sections instead.
> > This breaks the sk_page_frag() heuristic since the allocation
> > constraints are now stored in current->flags, which sk_page_frag()
> > can't read without risking triggering a cache miss and slowing down
> > TCP's fast path.
> >
> > This patch creates a new field in struct sock, named sk_use_task_frag,
> > which sockets with memory reclaim constraints can set to false if they
> > can't safely use current->task_frag. In such cases, sk_page_frag() now
> > always returns the socket's page_frag (->sk_frag). The first user is
> > sunrpc, which needs to avoid using current->task_frag but can keep
> > ->sk_allocation set to GFP_KERNEL otherwise.
> >
> > Eventually, it might be possible to simplify sk_page_frag() by only
> > testing ->sk_use_task_frag and avoid relying on the ->sk_allocation
> > heuristic entirely (assuming other sockets will set ->sk_use_task_frag
> > according to their constraints in the future).
> >
> > The new ->sk_use_task_frag field is placed in a hole in struct sock and
> > belongs to a cache line shared with ->sk_shutdown. Therefore it should
> > be hot and shouldn't have negative performance impacts on TCP's fast
> > path (sk_shutdown is tested just before the while() loop in
> > tcp_sendmsg_locked()).
> >
> > Link: https://lore.kernel.org/netdev/b4d8cb09c913d3e34f853736f3f5628abfd7f4b6.1656699567.git.gnault@xxxxxxxxxx/
> > Signed-off-by: Guillaume Nault <gnault@xxxxxxxxxx>
> > ---
> >  include/net/sock.h | 11 +++++++++--
> >  net/core/sock.c    |  1 +
> >  2 files changed, 10 insertions(+), 2 deletions(-)
> >
> > diff --git a/include/net/sock.h b/include/net/sock.h
> > index d08cfe190a78..ffba9e95470d 100644
> > --- a/include/net/sock.h
> > +++ b/include/net/sock.h
> > @@ -318,6 +318,9 @@ struct sk_filter;
> >    *  @sk_stamp: time stamp of last packet received
> >    *  @sk_stamp_seq: lock for accessing sk_stamp on 32 bit architectures only
> >    *  @sk_tsflags: SO_TIMESTAMPING flags
> > +  *  @sk_use_task_frag: allow sk_page_frag() to use current->task_frag.
> > +                        Sockets that can be used under memory reclaim should
> > +                        set this to false.
> >    *  @sk_bind_phc: SO_TIMESTAMPING bind PHC index of PTP virtual clock
> >    *                for timestamping
> >    *  @sk_tskey: counter to disambiguate concurrent tstamp requests
> > @@ -504,6 +507,7 @@ struct sock {
> >  #endif
> >       u16                     sk_tsflags;
> >       u8                      sk_shutdown;
> > +     bool                    sk_use_task_frag;
> >       atomic_t                sk_tskey;
> >       atomic_t                sk_zckey;
>
> I think the above should be fine from a data locality PoV, as the used
> cacheline should be hot at sk_page_frag_refill() usage time, as
> sk_tsflags has been accessed just before.
>
> @Eric, does the above fit with the planned sock fields reordering?

Do not worry about that, this can be handled later if needed.

>
> Jakub noted we could use a bitfield here to be future proof for
> additional flags addition. I think in this specific case a bool is
> preferable, because we actually wont to discourage people to add more
> of such flags, and the search for holes (or the bool -> bitflag
> conversion) should give to such eventual future changes some additional
> thoughts.
>
> Thanks!
>
> Paolo
>