ice: xsk: Handle SW XDP ring wrap and bump tail more often
authorMaciej Fijalkowski <maciej.fijalkowski@intel.com>
Tue, 25 Jan 2022 16:04:41 +0000 (17:04 +0100)
committerDaniel Borkmann <daniel@iogearbox.net>
Thu, 27 Jan 2022 16:25:32 +0000 (17:25 +0100)
Currently, if ice_clean_rx_irq_zc() processed the whole ring and
next_to_use != 0, then ice_alloc_rx_buf_zc() would not refill the whole
ring even if the XSK buffer pool would have enough free entries (either
from fill ring or the internal recycle mechanism) - it is because ring
wrap is not handled.

Improve the logic in ice_alloc_rx_buf_zc() to address the problem above.
Do not clamp the count of buffers that is passed to
xsk_buff_alloc_batch() in case when next_to_use + buffer count >=
rx_ring->count,  but rather split it and have two calls to the mentioned
function - one for the part up until the wrap and one for the part after
the wrap.

Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Alexander Lobakin <alexandr.lobakin@intel.com>
Link: https://lore.kernel.org/bpf/20220125160446.78976-4-maciej.fijalkowski@intel.com
drivers/net/ethernet/intel/ice/ice_txrx.h
drivers/net/ethernet/intel/ice/ice_xsk.c

index b7b3bd4..f70a5eb 100644 (file)
@@ -111,6 +111,8 @@ static inline int ice_skb_pad(void)
        (u16)((((R)->next_to_clean > (R)->next_to_use) ? 0 : (R)->count) + \
              (R)->next_to_clean - (R)->next_to_use - 1)
 
+#define ICE_RING_QUARTER(R) ((R)->count >> 2)
+
 #define ICE_TX_FLAGS_TSO       BIT(0)
 #define ICE_TX_FLAGS_HW_VLAN   BIT(1)
 #define ICE_TX_FLAGS_SW_VLAN   BIT(2)
index 27b0878..b12079a 100644 (file)
@@ -367,33 +367,28 @@ failure:
 }
 
 /**
- * ice_alloc_rx_bufs_zc - allocate a number of Rx buffers
- * @rx_ring: Rx ring
+ * ice_fill_rx_descs - pick buffers from XSK buffer pool and use it
+ * @pool: XSK Buffer pool to pull the buffers from
+ * @xdp: SW ring of xdp_buff that will hold the buffers
+ * @rx_desc: Pointer to Rx descriptors that will be filled
  * @count: The number of buffers to allocate
  *
  * This function allocates a number of Rx buffers from the fill ring
  * or the internal recycle mechanism and places them on the Rx ring.
  *
- * Returns true if all allocations were successful, false if any fail.
+ * Note that ring wrap should be handled by caller of this function.
+ *
+ * Returns the amount of allocated Rx descriptors
  */
-bool ice_alloc_rx_bufs_zc(struct ice_rx_ring *rx_ring, u16 count)
+static u16 ice_fill_rx_descs(struct xsk_buff_pool *pool, struct xdp_buff **xdp,
+                            union ice_32b_rx_flex_desc *rx_desc, u16 count)
 {
-       union ice_32b_rx_flex_desc *rx_desc;
-       u16 ntu = rx_ring->next_to_use;
-       struct xdp_buff **xdp;
-       u32 nb_buffs, i;
        dma_addr_t dma;
+       u16 buffs;
+       int i;
 
-       rx_desc = ICE_RX_DESC(rx_ring, ntu);
-       xdp = ice_xdp_buf(rx_ring, ntu);
-
-       nb_buffs = min_t(u16, count, rx_ring->count - ntu);
-       nb_buffs = xsk_buff_alloc_batch(rx_ring->xsk_pool, xdp, nb_buffs);
-       if (!nb_buffs)
-               return false;
-
-       i = nb_buffs;
-       while (i--) {
+       buffs = xsk_buff_alloc_batch(pool, xdp, count);
+       for (i = 0; i < buffs; i++) {
                dma = xsk_buff_xdp_get_dma(*xdp);
                rx_desc->read.pkt_addr = cpu_to_le64(dma);
                rx_desc->wb.status_error0 = 0;
@@ -402,13 +397,77 @@ bool ice_alloc_rx_bufs_zc(struct ice_rx_ring *rx_ring, u16 count)
                xdp++;
        }
 
+       return buffs;
+}
+
+/**
+ * __ice_alloc_rx_bufs_zc - allocate a number of Rx buffers
+ * @rx_ring: Rx ring
+ * @count: The number of buffers to allocate
+ *
+ * Place the @count of descriptors onto Rx ring. Handle the ring wrap
+ * for case where space from next_to_use up to the end of ring is less
+ * than @count. Finally do a tail bump.
+ *
+ * Returns true if all allocations were successful, false if any fail.
+ */
+static bool __ice_alloc_rx_bufs_zc(struct ice_rx_ring *rx_ring, u16 count)
+{
+       union ice_32b_rx_flex_desc *rx_desc;
+       u32 nb_buffs_extra = 0, nb_buffs;
+       u16 ntu = rx_ring->next_to_use;
+       u16 total_count = count;
+       struct xdp_buff **xdp;
+
+       rx_desc = ICE_RX_DESC(rx_ring, ntu);
+       xdp = ice_xdp_buf(rx_ring, ntu);
+
+       if (ntu + count >= rx_ring->count) {
+               nb_buffs_extra = ice_fill_rx_descs(rx_ring->xsk_pool, xdp,
+                                                  rx_desc,
+                                                  rx_ring->count - ntu);
+               rx_desc = ICE_RX_DESC(rx_ring, 0);
+               xdp = ice_xdp_buf(rx_ring, 0);
+               ntu = 0;
+               count -= nb_buffs_extra;
+               ice_release_rx_desc(rx_ring, 0);
+       }
+
+       nb_buffs = ice_fill_rx_descs(rx_ring->xsk_pool, xdp, rx_desc, count);
+
        ntu += nb_buffs;
        if (ntu == rx_ring->count)
                ntu = 0;
 
-       ice_release_rx_desc(rx_ring, ntu);
+       if (rx_ring->next_to_use != ntu)
+               ice_release_rx_desc(rx_ring, ntu);
+
+       return total_count == (nb_buffs_extra + nb_buffs);
+}
+
+/**
+ * ice_alloc_rx_bufs_zc - allocate a number of Rx buffers
+ * @rx_ring: Rx ring
+ * @count: The number of buffers to allocate
+ *
+ * Wrapper for internal allocation routine; figure out how many tail
+ * bumps should take place based on the given threshold
+ *
+ * Returns true if all calls to internal alloc routine succeeded
+ */
+bool ice_alloc_rx_bufs_zc(struct ice_rx_ring *rx_ring, u16 count)
+{
+       u16 rx_thresh = ICE_RING_QUARTER(rx_ring);
+       u16 batched, leftover, i, tail_bumps;
+
+       batched = ALIGN_DOWN(count, rx_thresh);
+       tail_bumps = batched / rx_thresh;
+       leftover = count & (rx_thresh - 1);
 
-       return count == nb_buffs;
+       for (i = 0; i < tail_bumps; i++)
+               if (!__ice_alloc_rx_bufs_zc(rx_ring, rx_thresh))
+                       return false;
+       return __ice_alloc_rx_bufs_zc(rx_ring, leftover);
 }
 
 /**