|
@@ -25,19 +25,8 @@
|
|
|
|
|
|
#include <linux/log2.h>
|
|
|
|
|
|
-/* slightly larger than one large A-MPDU */
|
|
|
-#define HTT_RX_RING_SIZE_MIN 128
|
|
|
-
|
|
|
-/* roughly 20 ms @ 1 Gbps of 1500B MSDUs */
|
|
|
-#define HTT_RX_RING_SIZE_MAX 2048
|
|
|
-
|
|
|
-#define HTT_RX_AVG_FRM_BYTES 1000
|
|
|
-
|
|
|
-/* ms, very conservative */
|
|
|
-#define HTT_RX_HOST_LATENCY_MAX_MS 20
|
|
|
-
|
|
|
-/* ms, conservative */
|
|
|
-#define HTT_RX_HOST_LATENCY_WORST_LIKELY_MS 10
|
|
|
+#define HTT_RX_RING_SIZE 1024
|
|
|
+#define HTT_RX_RING_FILL_LEVEL 1000
|
|
|
|
|
|
/* when under memory pressure rx ring refill may fail and needs a retry */
|
|
|
#define HTT_RX_RING_REFILL_RETRY_MS 50
|
|
@@ -45,68 +34,6 @@
|
|
|
static int ath10k_htt_rx_get_csum_state(struct sk_buff *skb);
|
|
|
static void ath10k_htt_txrx_compl_task(unsigned long ptr);
|
|
|
|
|
|
-static int ath10k_htt_rx_ring_size(struct ath10k_htt *htt)
|
|
|
-{
|
|
|
- int size;
|
|
|
-
|
|
|
- /*
|
|
|
- * It is expected that the host CPU will typically be able to
|
|
|
- * service the rx indication from one A-MPDU before the rx
|
|
|
- * indication from the subsequent A-MPDU happens, roughly 1-2 ms
|
|
|
- * later. However, the rx ring should be sized very conservatively,
|
|
|
- * to accomodate the worst reasonable delay before the host CPU
|
|
|
- * services a rx indication interrupt.
|
|
|
- *
|
|
|
- * The rx ring need not be kept full of empty buffers. In theory,
|
|
|
- * the htt host SW can dynamically track the low-water mark in the
|
|
|
- * rx ring, and dynamically adjust the level to which the rx ring
|
|
|
- * is filled with empty buffers, to dynamically meet the desired
|
|
|
- * low-water mark.
|
|
|
- *
|
|
|
- * In contrast, it's difficult to resize the rx ring itself, once
|
|
|
- * it's in use. Thus, the ring itself should be sized very
|
|
|
- * conservatively, while the degree to which the ring is filled
|
|
|
- * with empty buffers should be sized moderately conservatively.
|
|
|
- */
|
|
|
-
|
|
|
- /* 1e6 bps/mbps / 1e3 ms per sec = 1000 */
|
|
|
- size =
|
|
|
- htt->max_throughput_mbps +
|
|
|
- 1000 /
|
|
|
- (8 * HTT_RX_AVG_FRM_BYTES) * HTT_RX_HOST_LATENCY_MAX_MS;
|
|
|
-
|
|
|
- if (size < HTT_RX_RING_SIZE_MIN)
|
|
|
- size = HTT_RX_RING_SIZE_MIN;
|
|
|
-
|
|
|
- if (size > HTT_RX_RING_SIZE_MAX)
|
|
|
- size = HTT_RX_RING_SIZE_MAX;
|
|
|
-
|
|
|
- size = roundup_pow_of_two(size);
|
|
|
-
|
|
|
- return size;
|
|
|
-}
|
|
|
-
|
|
|
-static int ath10k_htt_rx_ring_fill_level(struct ath10k_htt *htt)
|
|
|
-{
|
|
|
- int size;
|
|
|
-
|
|
|
- /* 1e6 bps/mbps / 1e3 ms per sec = 1000 */
|
|
|
- size =
|
|
|
- htt->max_throughput_mbps *
|
|
|
- 1000 /
|
|
|
- (8 * HTT_RX_AVG_FRM_BYTES) * HTT_RX_HOST_LATENCY_WORST_LIKELY_MS;
|
|
|
-
|
|
|
- /*
|
|
|
- * Make sure the fill level is at least 1 less than the ring size.
|
|
|
- * Leaving 1 element empty allows the SW to easily distinguish
|
|
|
- * between a full ring vs. an empty ring.
|
|
|
- */
|
|
|
- if (size >= htt->rx_ring.size)
|
|
|
- size = htt->rx_ring.size - 1;
|
|
|
-
|
|
|
- return size;
|
|
|
-}
|
|
|
-
|
|
|
static void ath10k_htt_rx_ring_free(struct ath10k_htt *htt)
|
|
|
{
|
|
|
struct sk_buff *skb;
|
|
@@ -462,25 +389,18 @@ int ath10k_htt_rx_alloc(struct ath10k_htt *htt)
|
|
|
|
|
|
htt->rx_confused = false;
|
|
|
|
|
|
- htt->rx_ring.size = ath10k_htt_rx_ring_size(htt);
|
|
|
+ /* XXX: The fill level could be changed during runtime in response to
|
|
|
+ * the host processing latency. Is this really worth it?
|
|
|
+ */
|
|
|
+ htt->rx_ring.size = HTT_RX_RING_SIZE;
|
|
|
+ htt->rx_ring.size_mask = htt->rx_ring.size - 1;
|
|
|
+ htt->rx_ring.fill_level = HTT_RX_RING_FILL_LEVEL;
|
|
|
+
|
|
|
if (!is_power_of_2(htt->rx_ring.size)) {
|
|
|
ath10k_warn(ar, "htt rx ring size is not power of 2\n");
|
|
|
return -EINVAL;
|
|
|
}
|
|
|
|
|
|
- htt->rx_ring.size_mask = htt->rx_ring.size - 1;
|
|
|
-
|
|
|
- /*
|
|
|
- * Set the initial value for the level to which the rx ring
|
|
|
- * should be filled, based on the max throughput and the
|
|
|
- * worst likely latency for the host to fill the rx ring
|
|
|
- * with new buffers. In theory, this fill level can be
|
|
|
- * dynamically adjusted from the initial value set here, to
|
|
|
- * reflect the actual host latency rather than a
|
|
|
- * conservative assumption about the host latency.
|
|
|
- */
|
|
|
- htt->rx_ring.fill_level = ath10k_htt_rx_ring_fill_level(htt);
|
|
|
-
|
|
|
htt->rx_ring.netbufs_ring =
|
|
|
kzalloc(htt->rx_ring.size * sizeof(struct sk_buff *),
|
|
|
GFP_KERNEL);
|