Fixes invalid latency result when using the performance application and
hardware QAT PMD. It occurred when the number of processed packets was
higher then the size of the internal QAT PMD ring buffer and the buffer
was overflowed.
Fixed by correcting the registration of the enqueued packets and freeing
memory space for not enqueued packets.

Fixes: f8be1786b1b8 ("app/crypto-perf: introduce performance test application")

Signed-off-by: Slawomir Mrozowicz <slawomirx.mrozow...@intel.com>
---
 app/test-crypto-perf/cperf_test_latency.c | 20 ++++++++------------
 1 file changed, 8 insertions(+), 12 deletions(-)

diff --git a/app/test-crypto-perf/cperf_test_latency.c 
b/app/test-crypto-perf/cperf_test_latency.c
index 5ec1b2c..239a8e5 100644
--- a/app/test-crypto-perf/cperf_test_latency.c
+++ b/app/test-crypto-perf/cperf_test_latency.c
@@ -424,7 +424,6 @@ cperf_latency_test_runner(void *arg)
        struct rte_crypto_op *ops[ctx->options->burst_sz];
        struct rte_crypto_op *ops_processed[ctx->options->burst_sz];
        uint64_t ops_enqd = 0, ops_deqd = 0;
-       uint16_t ops_unused = 0;
        uint64_t m_idx = 0, b_idx = 0, i;
 
        uint64_t tsc_val, tsc_end, tsc_start;
@@ -460,19 +459,18 @@ cperf_latency_test_runner(void *arg)
                                                ctx->options->burst_sz :
                                                ctx->options->total_ops -
                                                enqd_tot;
-               uint16_t ops_needed = burst_size - ops_unused;
 
                /* Allocate crypto ops from pool */
-               if (ops_needed != rte_crypto_op_bulk_alloc(
+               if (burst_size != rte_crypto_op_bulk_alloc(
                                ctx->crypto_op_pool,
                                RTE_CRYPTO_OP_TYPE_SYMMETRIC,
-                               ops, ops_needed))
+                               ops, burst_size))
                        return -1;
 
                /* Setup crypto op, attach mbuf etc */
                (ctx->populate_ops)(ops, &ctx->mbufs_in[m_idx],
                                &ctx->mbufs_out[m_idx],
-                               ops_needed, ctx->sess, ctx->options,
+                               burst_size, ctx->sess, ctx->options,
                                ctx->test_vector);
 
                tsc_start = rte_rdtsc_precise();
@@ -498,17 +496,15 @@ cperf_latency_test_runner(void *arg)
 
                tsc_end = rte_rdtsc_precise();
 
-               for (i = 0; i < ops_needed; i++) {
+               for (i = 0; i < ops_enqd; i++) {
                        ctx->res[tsc_idx].tsc_start = tsc_start;
                        ops[i]->opaque_data = (void *)&ctx->res[tsc_idx];
                        tsc_idx++;
                }
 
-               /*
-                * Calculate number of ops not enqueued (mainly for hw
-                * accelerators whose ingress queue can fill up).
-                */
-               ops_unused = burst_size - ops_enqd;
+               /* Free memory for not enqueued operations */
+               for (i = ops_enqd; i < burst_size; i++)
+                       rte_crypto_op_free(ops[i]);
 
                if (likely(ops_deqd))  {
                        /*
@@ -535,7 +531,7 @@ cperf_latency_test_runner(void *arg)
                enqd_max = max(ops_enqd, enqd_max);
                enqd_min = min(ops_enqd, enqd_min);
 
-               m_idx += ops_needed;
+               m_idx += ops_enqd;
                m_idx = m_idx + ctx->options->burst_sz > ctx->options->pool_sz ?
                                0 : m_idx;
                b_idx++;
-- 
2.5.0

Reply via email to