Re: [next-queue v2 2/4] i40e: Record number TXes cleaned during NAPI

[Maciej Fijalkowski <maciej.fijalkowski@intel.com>]

On Wed, Oct 05, 2022 at 06:00:24PM -0700, Joe Damato wrote:
> On Wed, Oct 05, 2022 at 05:31:04PM -0700, Joe Damato wrote:
> > On Wed, Oct 05, 2022 at 07:16:56PM -0500, Samudrala, Sridhar wrote:
> > > On 10/5/2022 4:21 PM, Joe Damato wrote:
> > > >Update i40e_clean_tx_irq to take an out parameter (tx_cleaned) which stores
> > > >the number TXs cleaned.
> > > >
> > > >Likewise, update i40e_clean_xdp_tx_irq and i40e_xmit_zc to do the same.
> > > >
> > > >Care has been taken to avoid changing the control flow of any functions
> > > >involved.
> > > >
> > > >Signed-off-by: Joe Damato <jdamato@fastly.com>
> > > >---
> > > >  drivers/net/ethernet/intel/i40e/i40e_txrx.c | 16 +++++++++++-----
> > > >  drivers/net/ethernet/intel/i40e/i40e_xsk.c  | 15 +++++++++++----
> > > >  drivers/net/ethernet/intel/i40e/i40e_xsk.h  |  3 ++-
> > > >  3 files changed, 24 insertions(+), 10 deletions(-)
> > > >
> > > >diff --git a/drivers/net/ethernet/intel/i40e/i40e_txrx.c b/drivers/net/ethernet/intel/i40e/i40e_txrx.c
> > > >index b97c95f..a2cc98e 100644
> > > >--- a/drivers/net/ethernet/intel/i40e/i40e_txrx.c
> > > >+++ b/drivers/net/ethernet/intel/i40e/i40e_txrx.c
> > > >@@ -923,11 +923,13 @@ void i40e_detect_recover_hung(struct i40e_vsi *vsi)
> > > >   * @vsi: the VSI we care about
> > > >   * @tx_ring: Tx ring to clean
> > > >   * @napi_budget: Used to determine if we are in netpoll
> > > >+ * @tx_cleaned: Out parameter set to the number of TXes cleaned
> > > >   *
> > > >   * Returns true if there's any budget left (e.g. the clean is finished)
> > > >   **/
> > > >  static bool i40e_clean_tx_irq(struct i40e_vsi *vsi,
> > > >-			      struct i40e_ring *tx_ring, int napi_budget)
> > > >+			      struct i40e_ring *tx_ring, int napi_budget,
> > > >+			      unsigned int *tx_cleaned)
> > > >  {
> > > >  	int i = tx_ring->next_to_clean;
> > > >  	struct i40e_tx_buffer *tx_buf;
> > > >@@ -1026,7 +1028,7 @@ static bool i40e_clean_tx_irq(struct i40e_vsi *vsi,
> > > >  	i40e_arm_wb(tx_ring, vsi, budget);
> > > >  	if (ring_is_xdp(tx_ring))
> > > >-		return !!budget;
> > > >+		goto out;
> > > >  	/* notify netdev of completed buffers */
> > > >  	netdev_tx_completed_queue(txring_txq(tx_ring),
> > > >@@ -1048,6 +1050,8 @@ static bool i40e_clean_tx_irq(struct i40e_vsi *vsi,
> > > >  		}
> > > >  	}
> > > >+out:
> > > >+	*tx_cleaned = total_packets;
> > > >  	return !!budget;
> > > >  }
> > > >@@ -2689,10 +2693,12 @@ int i40e_napi_poll(struct napi_struct *napi, int budget)
> > > >  			       container_of(napi, struct i40e_q_vector, napi);
> > > >  	struct i40e_vsi *vsi = q_vector->vsi;
> > > >  	struct i40e_ring *ring;
> > > >+	bool tx_clean_complete = true;
> > > >  	bool clean_complete = true;
> > > >  	bool arm_wb = false;
> > > >  	int budget_per_ring;
> > > >  	int work_done = 0;
> > > >+	unsigned int tx_cleaned = 0;
> > > >  	if (test_bit(__I40E_VSI_DOWN, vsi->state)) {
> > > >  		napi_complete(napi);
> > > >@@ -2704,11 +2710,11 @@ int i40e_napi_poll(struct napi_struct *napi, int budget)
> > > >  	 */
> > > >  	i40e_for_each_ring(ring, q_vector->tx) {
> > > >  		bool wd = ring->xsk_pool ?
> > > >-			  i40e_clean_xdp_tx_irq(vsi, ring) :
> > > >-			  i40e_clean_tx_irq(vsi, ring, budget);
> > > >+			  i40e_clean_xdp_tx_irq(vsi, ring, &tx_cleaned) :
> > > >+			  i40e_clean_tx_irq(vsi, ring, budget, &tx_cleaned);
> > > >  		if (!wd) {
> > > >-			clean_complete = false;
> > > >+			clean_complete = tx_clean_complete = false;
> > > >  			continue;
> > > >  		}
> > > >  		arm_wb |= ring->arm_wb;
> > > >diff --git a/drivers/net/ethernet/intel/i40e/i40e_xsk.c b/drivers/net/ethernet/intel/i40e/i40e_xsk.c
> > > >index 790aaeff..f98ce7e4 100644
> > > >--- a/drivers/net/ethernet/intel/i40e/i40e_xsk.c
> > > >+++ b/drivers/net/ethernet/intel/i40e/i40e_xsk.c
> > > >@@ -530,18 +530,22 @@ static void i40e_set_rs_bit(struct i40e_ring *xdp_ring)
> > > >   * i40e_xmit_zc - Performs zero-copy Tx AF_XDP
> > > >   * @xdp_ring: XDP Tx ring
> > > >   * @budget: NAPI budget
> > > >+ * @tx_cleaned: Out parameter of the TX packets processed
> > > >   *
> > > >   * Returns true if the work is finished.
> > > >   **/
> > > >-static bool i40e_xmit_zc(struct i40e_ring *xdp_ring, unsigned int budget)
> > > >+static bool i40e_xmit_zc(struct i40e_ring *xdp_ring, unsigned int budget,
> > > >+			 unsigned int *tx_cleaned)
> > > >  {
> > > >  	struct xdp_desc *descs = xdp_ring->xsk_pool->tx_descs;
> > > >  	u32 nb_pkts, nb_processed = 0;
> > > >  	unsigned int total_bytes = 0;
> > > >  	nb_pkts = xsk_tx_peek_release_desc_batch(xdp_ring->xsk_pool, budget);
> > > >-	if (!nb_pkts)
> > > >+	if (!nb_pkts) {
> > > >+		*tx_cleaned = 0;
> > > >  		return true;
> > > >+	}
> > > >  	if (xdp_ring->next_to_use + nb_pkts >= xdp_ring->count) {
> > > >  		nb_processed = xdp_ring->count - xdp_ring->next_to_use;
> > > >@@ -558,6 +562,7 @@ static bool i40e_xmit_zc(struct i40e_ring *xdp_ring, unsigned int budget)
> > > >  	i40e_update_tx_stats(xdp_ring, nb_pkts, total_bytes);
> > > >+	*tx_cleaned = nb_pkts;
> > > 
> > > With XDP, I don't think we should count these as tx_cleaned packets. These are transmitted
> > > packets. The tx_cleaned would be the xsk_frames counter in i40e_clean_xdp_tx_irq
> > > May be we need 2 counters for xdp.
> > 
> > I think there's two issues you are describing, which are separate in my
> > mind.
> > 
> >   1.) The name "tx_cleaned", and
> >   2.) Whether nb_pkts is the right thing to write as the out param.
> > 
> > For #1: I'm OK to change the name if that's the blocker here; please
> > suggest a suitable alternative that you'll accept.
> > 
> > For #2: nb_pkts is, IMO, the right value to bubble up to the tracepoint because
> > nb_pkts affects clean_complete in i40e_napi_poll which in turn determines
> > whether or not polling mode is entered.
> > 
> > The purpose of the tracepoint is to determine when/why/how you are entering
> > polling mode, so if nb_pkts plays a role in that calculation, it's the
> > right number to output.
> 
> I suppose the alternative is to only fire the tracepoint when *not* in XDP.
> Then the changes to the XDP stuff can be dropped and a separate set of
> tracepoints for XDP can be created in the future.

Let's be clear that it's the AF_XDP quirk that we have in here that actual
xmit happens within NAPI polling routine.

Sridhar is right with having xsk_frames as tx_cleaned but you're also
right that nb_pkts affects napi polling. But then if you look at Rx side
there is an analogous case with buffer allocation affecting napi polling.

> 
> That might reduce the complexity a bit, and will probably still be pretty
> useful for people tuning their non-XDP workloads.