diff options
author | Steven Luong <sluong@cisco.com> | 2020-02-13 09:27:38 -0800 |
---|---|---|
committer | Dave Barach <openvpp@barachs.net> | 2020-02-25 22:37:44 +0000 |
commit | f7ab7b2d9bc0f42c1e766d22d49dd0dc4f28abb6 (patch) | |
tree | 4bbea14cfb291fb5754a87c1575e65edd68d704c /src | |
parent | 2da39718f560478678caacccd198ee4c0c9673c3 (diff) |
avf: Handle chain buffer in TX properly
For chain buffer, need to traverse b->next_buffer to transmit all buffers
in the chain. Only set EOP at the last descriptor in the chain to signal
this is a chain descriptor.
Introduce slow path to handle ring wrap. This is needed because chain
buffer may consist of multiple pieces and it may span from near the end of
the ring to the beginning of the ring.
Type: fix
Signed-off-by: Steven Luong <sluong@cisco.com>
Change-Id: Id7c872f3e39e09f3566aa63f3cdba8f40736d508
Diffstat (limited to 'src')
-rw-r--r-- | src/plugins/avf/avf.h | 3 | ||||
-rw-r--r-- | src/plugins/avf/output.c | 144 |
2 files changed, 140 insertions, 7 deletions
diff --git a/src/plugins/avf/avf.h b/src/plugins/avf/avf.h index a8e8b96e18b..744b58370e4 100644 --- a/src/plugins/avf/avf.h +++ b/src/plugins/avf/avf.h @@ -333,7 +333,8 @@ typedef struct } avf_input_trace_t; #define foreach_avf_tx_func_error \ -_(NO_FREE_SLOTS, "no free tx slots") + _(SEGMENT_SIZE_EXCEEDED, "segment size exceeded") \ + _(NO_FREE_SLOTS, "no free tx slots") typedef enum { diff --git a/src/plugins/avf/output.c b/src/plugins/avf/output.c index 1e38377e844..6c43885569e 100644 --- a/src/plugins/avf/output.c +++ b/src/plugins/avf/output.c @@ -31,8 +31,8 @@ avf_tx_desc_get_dtyp (avf_tx_desc_t * d) } static_always_inline u16 -avf_tx_enqueue (vlib_main_t * vm, avf_txq_t * txq, u32 * buffers, - u32 n_packets, int use_va_dma) +avf_tx_enqueue (vlib_main_t * vm, vlib_node_runtime_t * node, avf_txq_t * txq, + u32 * buffers, u32 n_packets, int use_va_dma) { u16 next = txq->next; u64 bits = AVF_TXD_CMD_EOP | AVF_TXD_CMD_RSV; @@ -41,6 +41,8 @@ avf_tx_enqueue (vlib_main_t * vm, avf_txq_t * txq, u32 * buffers, u16 mask = txq->size - 1; vlib_buffer_t *b[4]; avf_tx_desc_t *d = txq->descs + next; + u16 n_desc_needed; + vlib_buffer_t *b0; /* avoid ring wrap */ n_desc_left = txq->size - clib_max (txq->next, txq->n_enqueued + 8); @@ -48,6 +50,7 @@ avf_tx_enqueue (vlib_main_t * vm, avf_txq_t * txq, u32 * buffers, if (n_desc_left == 0) return 0; + /* Fast path, no ring wrap */ while (n_packets_left && n_desc_left) { u32 or_flags; @@ -103,6 +106,57 @@ avf_tx_enqueue (vlib_main_t * vm, avf_txq_t * txq, u32 * buffers, txq->bufs[next] = buffers[0]; b[0] = vlib_get_buffer (vm, buffers[0]); + /* Deal with chain buffer if present */ + if (b[0]->flags & VLIB_BUFFER_NEXT_PRESENT) + { + n_desc_needed = 1; + b0 = b[0]; + + /* Wish there were a buffer count for chain buffer */ + while (b0->flags & VLIB_BUFFER_NEXT_PRESENT) + { + b0 = vlib_get_buffer (vm, b0->next_buffer); + n_desc_needed++; + } + + /* spec says data descriptor is limited to 8 segments */ + if (PREDICT_FALSE (n_desc_needed > 8)) + { + vlib_buffer_free_one (vm, buffers[0]); + vlib_error_count (vm, node->node_index, + AVF_TX_ERROR_SEGMENT_SIZE_EXCEEDED, 1); + n_packets_left -= 1; + buffers += 1; + continue; + } + + if (PREDICT_FALSE (n_desc_left < n_desc_needed)) + /* + * Slow path may be able to to deal with this since it can handle + * ring wrap + */ + break; + + while (b[0]->flags & VLIB_BUFFER_NEXT_PRESENT) + { + if (use_va_dma) + d[0].qword[0] = vlib_buffer_get_current_va (b[0]); + else + d[0].qword[0] = vlib_buffer_get_current_pa (vm, b[0]); + + d[0].qword[1] = (((u64) b[0]->current_length) << 34) | + AVF_TXD_CMD_RSV; + + next += 1; + n_desc += 1; + n_desc_left -= 1; + d += 1; + + txq->bufs[next] = b[0]->next_buffer; + b[0] = vlib_get_buffer (vm, b[0]->next_buffer); + } + } + if (use_va_dma) d[0].qword[0] = vlib_buffer_get_current_va (b[0]); else @@ -118,6 +172,84 @@ avf_tx_enqueue (vlib_main_t * vm, avf_txq_t * txq, u32 * buffers, d += 1; } + /* Slow path to support ring wrap */ + if (PREDICT_FALSE (n_packets_left)) + { + txq->n_enqueued += n_desc; + + n_desc = 0; + d = txq->descs + (next & mask); + + /* +8 to be consistent with fast path */ + n_desc_left = txq->size - (txq->n_enqueued + 8); + + while (n_packets_left && n_desc_left) + { + txq->bufs[next & mask] = buffers[0]; + b[0] = vlib_get_buffer (vm, buffers[0]); + + /* Deal with chain buffer if present */ + if (b[0]->flags & VLIB_BUFFER_NEXT_PRESENT) + { + n_desc_needed = 1; + b0 = b[0]; + + while (b0->flags & VLIB_BUFFER_NEXT_PRESENT) + { + b0 = vlib_get_buffer (vm, b0->next_buffer); + n_desc_needed++; + } + + /* Spec says data descriptor is limited to 8 segments */ + if (PREDICT_FALSE (n_desc_needed > 8)) + { + vlib_buffer_free_one (vm, buffers[0]); + vlib_error_count (vm, node->node_index, + AVF_TX_ERROR_SEGMENT_SIZE_EXCEEDED, 1); + n_packets_left -= 1; + buffers += 1; + continue; + } + + if (PREDICT_FALSE (n_desc_left < n_desc_needed)) + break; + + while (b[0]->flags & VLIB_BUFFER_NEXT_PRESENT) + { + if (use_va_dma) + d[0].qword[0] = vlib_buffer_get_current_va (b[0]); + else + d[0].qword[0] = vlib_buffer_get_current_pa (vm, b[0]); + + d[0].qword[1] = (((u64) b[0]->current_length) << 34) | + AVF_TXD_CMD_RSV; + + next += 1; + n_desc += 1; + n_desc_left -= 1; + d = txq->descs + (next & mask); + + txq->bufs[next & mask] = b[0]->next_buffer; + b[0] = vlib_get_buffer (vm, b[0]->next_buffer); + } + } + + if (use_va_dma) + d[0].qword[0] = vlib_buffer_get_current_va (b[0]); + else + d[0].qword[0] = vlib_buffer_get_current_pa (vm, b[0]); + + d[0].qword[1] = (((u64) b[0]->current_length) << 34) | bits; + + next += 1; + n_desc += 1; + buffers += 1; + n_packets_left -= 1; + n_desc_left -= 1; + d = txq->descs + (next & mask); + } + } + if ((slot = clib_ring_enq (txq->rs_slots))) { u16 rs_slot = slot[0] = (next - 1) & mask; @@ -177,15 +309,15 @@ retry: n_free = (complete_slot + 1 - first) & mask; txq->n_enqueued -= n_free; - vlib_buffer_free_from_ring (vm, txq->bufs, first, txq->size, - n_free); + vlib_buffer_free_from_ring_no_next (vm, txq->bufs, first, txq->size, + n_free); } } if (ad->flags & AVF_DEVICE_F_VA_DMA) - n_enq = avf_tx_enqueue (vm, txq, buffers, n_left, 1); + n_enq = avf_tx_enqueue (vm, node, txq, buffers, n_left, 1); else - n_enq = avf_tx_enqueue (vm, txq, buffers, n_left, 0); + n_enq = avf_tx_enqueue (vm, node, txq, buffers, n_left, 0); n_left -= n_enq; |