On Wed, Jul 15, 2020 at 4:21 PM Jakub Kicinski <k...@kernel.org> wrote:
>
> On Tue, 7 Jul 2020 17:28:04 +0206 John Ogness wrote:
> > A busy-wait loop is used to implement waiting for bits to be copied
> > from the skb to the kernel buffer before retiring a block. This is
> > a problem on PREEMPT_RT because the copying task could be preempted
> > by the busy-waiting task and thus live lock in the busy-wait loop.
> >
> > Replace the busy-wait logic with an rwlock_t. This provides lockdep
> > coverage and makes the code RT ready.
> >
> > Signed-off-by: John Ogness <john.ogn...@linutronix.de>
>
> Is taking a lock and immediately releasing it better than a completion?
> Seems like the lock is guaranteed to dirty a cache line, which would
> otherwise be avoided here.
>
> Willem, would you be able to take a look as well? Is this path
> performance sensitive in real life?
No objections from me.
I guess this resolves the issue on preempt_rt, because the spinlocks act as
mutexes. It will still spin on write_lock otherwise, no huge difference from
existing logic.
>
> > diff --git a/net/packet/af_packet.c b/net/packet/af_packet.c
> > index 29bd405adbbd..dd1eec2dd6ef 100644
> > --- a/net/packet/af_packet.c
> > +++ b/net/packet/af_packet.c
> > @@ -593,6 +593,7 @@ static void init_prb_bdqc(struct packet_sock *po,
> > req_u->req3.tp_block_size);
> > p1->tov_in_jiffies = msecs_to_jiffies(p1->retire_blk_tov);
> > p1->blk_sizeof_priv = req_u->req3.tp_sizeof_priv;
> > + rwlock_init(&p1->blk_fill_in_prog_lock);
> >
> > p1->max_frame_len = p1->kblk_size -
> > BLK_PLUS_PRIV(p1->blk_sizeof_priv);
> > prb_init_ft_ops(p1, req_u);
> > @@ -659,10 +660,9 @@ static void prb_retire_rx_blk_timer_expired(struct
> > timer_list *t)
> > *
> > */
> > if (BLOCK_NUM_PKTS(pbd)) {
> > - while (atomic_read(&pkc->blk_fill_in_prog)) {
> > - /* Waiting for skb_copy_bits to finish... */
> > - cpu_relax();
> > - }
> > + /* Waiting for skb_copy_bits to finish... */
> > + write_lock(&pkc->blk_fill_in_prog_lock);
> > + write_unlock(&pkc->blk_fill_in_prog_lock);
> > }
> >
> > if (pkc->last_kactive_blk_num == pkc->kactive_blk_num) {
> > @@ -921,10 +921,9 @@ static void prb_retire_current_block(struct
> > tpacket_kbdq_core *pkc,
> > * the timer-handler already handled this case.
> > */
> > if (!(status & TP_STATUS_BLK_TMO)) {
> > - while (atomic_read(&pkc->blk_fill_in_prog)) {
> > - /* Waiting for skb_copy_bits to finish... */
> > - cpu_relax();
> > - }
> > + /* Waiting for skb_copy_bits to finish... */
> > + write_lock(&pkc->blk_fill_in_prog_lock);
> > + write_unlock(&pkc->blk_fill_in_prog_lock);
> > }
> > prb_close_block(pkc, pbd, po, status);
> > return;
> > @@ -944,7 +943,8 @@ static int prb_queue_frozen(struct tpacket_kbdq_core
> > *pkc)
> > static void prb_clear_blk_fill_status(struct packet_ring_buffer *rb)
> > {
> > struct tpacket_kbdq_core *pkc = GET_PBDQC_FROM_RB(rb);
> > - atomic_dec(&pkc->blk_fill_in_prog);
> > +
> > + read_unlock(&pkc->blk_fill_in_prog_lock);
> > }
> >
> > static void prb_fill_rxhash(struct tpacket_kbdq_core *pkc,
> > @@ -998,7 +998,7 @@ static void prb_fill_curr_block(char *curr,
> > pkc->nxt_offset += TOTAL_PKT_LEN_INCL_ALIGN(len);
> > BLOCK_LEN(pbd) += TOTAL_PKT_LEN_INCL_ALIGN(len);
> > BLOCK_NUM_PKTS(pbd) += 1;
> > - atomic_inc(&pkc->blk_fill_in_prog);
> > + read_lock(&pkc->blk_fill_in_prog_lock);
> > prb_run_all_ft_ops(pkc, ppd);
> > }
> >
> > diff --git a/net/packet/internal.h b/net/packet/internal.h
> > index 907f4cd2a718..fd41ecb7f605 100644
> > --- a/net/packet/internal.h
> > +++ b/net/packet/internal.h
> > @@ -39,7 +39,7 @@ struct tpacket_kbdq_core {
> > char *nxt_offset;
> > struct sk_buff *skb;
> >
> > - atomic_t blk_fill_in_prog;
> > + rwlock_t blk_fill_in_prog_lock;
> >
> > /* Default is set to 8ms */
> > #define DEFAULT_PRB_RETIRE_TOV (8)
>
