[v2,3/3] net/ark: support chunk DMA transfers
Checks
Commit Message
---
v2:
Various performance optimizations and behavior fixes
- Chunk mpu transfer use 64 objects (512 byte) to maintain memory
read alignment
- Align mpu memory allocation to be at 512 byte boundaries
- Reduce force-close allocation from 10000 objects to 64 objects
- Add memory write barriers for read and wait status functions
in ddm, udm and mpu.
- Configuration status updates for internal packet checker and generator.
---
Signed-off-by: John Miller <john.miller@atomicrules.com>
---
drivers/net/ark/ark_ddm.c | 1 +
drivers/net/ark/ark_ethdev_rx.c | 16 +++++++++-------
drivers/net/ark/ark_mpu.c | 1 +
drivers/net/ark/ark_pktchkr.c | 2 +-
drivers/net/ark/ark_pktgen.c | 2 +-
drivers/net/ark/ark_udm.c | 3 +++
6 files changed, 16 insertions(+), 9 deletions(-)
@@ -55,6 +55,7 @@ ark_ddm_stop(struct ark_ddm_t *ddm, const int wait)
int cnt = 0;
ddm->cfg.command = 2;
+ rte_wmb();
while (wait && (ddm->cfg.stop_flushed & 0x01) == 0) {
if (cnt++ > 1000)
return 1;
@@ -12,6 +12,7 @@
#define ARK_RX_META_SIZE 32
#define ARK_RX_META_OFFSET (RTE_PKTMBUF_HEADROOM - ARK_RX_META_SIZE)
+#define ARK_RX_MPU_CHUNK (64U)
/* Forward declarations */
struct ark_rx_queue;
@@ -104,7 +105,7 @@ static inline void
eth_ark_rx_update_cons_index(struct ark_rx_queue *queue, uint32_t cons_index)
{
queue->cons_index = cons_index;
- if ((cons_index + queue->queue_size - queue->seed_index) >= 64U) {
+ if ((cons_index + queue->queue_size - queue->seed_index) >= ARK_RX_MPU_CHUNK) {
eth_ark_rx_seed_mbufs(queue);
ark_mpu_set_producer(queue->mpu, queue->seed_index);
}
@@ -179,12 +180,12 @@ eth_ark_dev_rx_queue_setup(struct rte_eth_dev *dev,
queue->reserve_q =
rte_zmalloc_socket("Ark_rx_queue mbuf",
nb_desc * sizeof(struct rte_mbuf *),
- 64,
+ 512,
socket_id);
queue->paddress_q =
rte_zmalloc_socket("Ark_rx_queue paddr",
nb_desc * sizeof(rte_iova_t),
- 64,
+ 512,
socket_id);
if (queue->reserve_q == 0 || queue->paddress_q == 0) {
@@ -455,7 +456,8 @@ eth_ark_rx_stop_queue(struct rte_eth_dev *dev, uint16_t queue_id)
static inline int
eth_ark_rx_seed_mbufs(struct ark_rx_queue *queue)
{
- uint32_t limit = queue->cons_index + queue->queue_size;
+ uint32_t limit = (queue->cons_index & ~(ARK_RX_MPU_CHUNK - 1)) +
+ queue->queue_size;
uint32_t seed_index = queue->seed_index;
uint32_t count = 0;
@@ -618,14 +620,14 @@ eth_ark_udm_force_close(struct rte_eth_dev *dev)
ark_mpu_start(queue->mpu);
/* Add some buffers */
- index = 100000 + queue->seed_index;
+ index = ARK_RX_MPU_CHUNK + queue->seed_index;
ark_mpu_set_producer(queue->mpu, index);
}
/* Wait to allow data to pass */
usleep(100);
- ARK_PMD_LOG(DEBUG, "UDM forced flush attempt, stopped = %d\n",
- ark_udm_is_flushed(ark->udm.v));
+ ARK_PMD_LOG(NOTICE, "UDM forced flush attempt, stopped = %d\n",
+ ark_udm_is_flushed(ark->udm.v));
}
ark_udm_reset(ark->udm.v);
}
@@ -68,6 +68,7 @@ ark_mpu_reset(struct ark_mpu_t *mpu)
int cnt = 0;
mpu->cfg.command = MPU_CMD_RESET;
+ rte_wmb();
while (mpu->cfg.command != MPU_CMD_IDLE) {
if (cnt++ > 1000)
@@ -113,7 +113,7 @@ ark_pktchkr_stopped(ark_pkt_chkr_t handle)
struct ark_pkt_chkr_inst *inst = (struct ark_pkt_chkr_inst *)handle;
uint32_t r = inst->sregs->pkt_start_stop;
- return (((r >> 16) & 1) == 1);
+ return (((r >> 16) & 1) == 1) || (r == 0);
}
void
@@ -107,7 +107,7 @@ ark_pktgen_paused(ark_pkt_gen_t handle)
struct ark_pkt_gen_inst *inst = (struct ark_pkt_gen_inst *)handle;
uint32_t r = inst->regs->pkt_start_stop;
- return (((r >> 16) & 1) == 1);
+ return (((r >> 24) & 1) == 1) || (((r >> 16) & 1) == 1) || (r == 0);
}
void
@@ -33,7 +33,9 @@ ark_udm_stop(struct ark_udm_t *udm, const int wait)
{
int cnt = 0;
+ udm->setup.r0 = 0;
udm->cfg.command = 2;
+ rte_wmb();
while (wait && (udm->cfg.stop_flushed & 0x01) == 0) {
if (cnt++ > 1000)
@@ -70,6 +72,7 @@ ark_udm_reset(struct ark_udm_t *udm)
void
ark_udm_start(struct ark_udm_t *udm)
{
+ udm->setup.r0 = 0x100;
udm->cfg.command = 1;
}