On Sat, Mar 2, 2019 at 11:41 AM Pavel Shilovsky <piastryyy@xxxxxxxxx> wrote:
>
> ср, 27 февр. 2019 г. в 22:26, Ronnie Sahlberg <lsahlber@xxxxxxxxxx>:
> >
> > Reserve the last MAX_COMPOUND credits for any request asking for >1 credit.
> > This is to prevent future compound requests from becoming starved while waiting
> > for potentially many requests is there is a large number of concurrent
> > singe-credit requests.
> >
> > Signed-off-by: Ronnie Sahlberg <lsahlber@xxxxxxxxxx>
> > ---
> > fs/cifs/transport.c | 23 +++++++++++++++++++++++
> > 1 file changed, 23 insertions(+)
> >
> > diff --git a/fs/cifs/transport.c b/fs/cifs/transport.c
> > index baf15194aa3d..4ff832fce2e9 100644
> > --- a/fs/cifs/transport.c
> > +++ b/fs/cifs/transport.c
> > @@ -520,6 +520,29 @@ wait_for_free_credits(struct TCP_Server_Info *server, const int num_credits,
> > }
> >
> > /*
> > + * For normal commands, reserve the last MAX_COMPOUND
> > + * credits to compound requests.
> > + * Otherwise these compounds could be permanently
> > + * starved for credits by single-credit requests.
> > + *
> > + * To prevent spinning CPU, block this thread until
> > + * there are >MAX_COMPOUND credits available.
> > + */
> > + if (!optype && num_credits == 1 &&
>
> What if server decides to grants us credits gradually? E.g. granted
> only 2 during negotiate protocol and session setup. The tree connect
> will wait forever here.
Thanks. You are right.
That should be avoidable by checking for and only doing this if we
have <arbitrary large> number of credits already in flight.
I will send an updated series.
>
> > + *credits <= MAX_COMPOUND) {
> > + spin_unlock(&server->req_lock);
> > + cifs_num_waiters_inc(server);
> > + rc = wait_event_killable(server->request_q,
> > + has_credits(server, credits,
> > + MAX_COMPOUND + 1));
> > + cifs_num_waiters_dec(server);
> > + if (rc)
> > + return rc;
> > + spin_lock(&server->req_lock);
> > + continue;
> > + }
> > +
> > + /*
> > * Can not count locking commands against total
> > * as they are allowed to block on server.
> > */
> > --
> > 2.13.6
> >
>
>
> --
> Best regards,
> Pavel Shilovsky
