ath11k: drop tx_status_fifo
RFC 1/2 ath11k: drop tx_status_fifo
perf top showed that 10% of the cpu time was spent on the memcpy used to populate the fifo when the device is under high load. There is no apparent benefit from using a fifo so lets just drop it.
--- drivers/net/wireless/ath/ath11k/dp.c | 13 --------- drivers/net/wireless/ath/ath11k/dp.h | 5 ---- drivers/net/wireless/ath/ath11k/dp_tx.c | 37 +++++------------------- drivers/net/wireless/ath/ath11k/hal_tx.c | 5 ---- drivers/net/wireless/ath/ath11k/hal_tx.h | 1 - 5 files changed, 8 insertions(+), 53 deletions(-)
drivers/net/wireless/ath/ath11k/dp.c +0 -13
@@ -839,10 +839,6 @@ void ath11k_dp_free(struct ath11k_base *sc) ath11k_dp_tx_pending_cleanup, sc);idr_destroy (&dp->tx_ring[i].txbuf_idr);spin_unlock_bh (&dp->tx_ring[i].tx_idr_lock); - -spin_lock_bh (&dp->tx_ring[i].tx_status_lock); -kfifo_free (&dp->tx_ring[i].tx_status_fifo); -spin_unlock_bh (&dp->tx_ring[i].tx_status_lock); }/* Deinit any SOC level resource */
@@ -888,12 +884,6 @@ int ath11k_dp_alloc(struct ath11k_base *sc)idr_init (&dp->tx_ring[i].txbuf_idr);spin_lock_init (&dp->tx_ring[i].tx_idr_lock); dp->tx_ring[i].tcl_data_ring_id = i; - -spin_lock_init (&dp->tx_ring[i].tx_status_lock); - ret =kfifo_alloc (&dp->tx_ring[i].tx_status_fifo, size, -GFP_KERNEL ); -if (ret) - goto fail_cmn_srng_cleanup; }for (i =0 ; i <HAL_DSCP_TID_MAP_TBL_NUM_ENTRIES_MAX ; i++)
@@ -903,9 +893,6 @@ int ath11k_dp_alloc(struct ath11k_base *sc)return 0 ; -fail_cmn_srng_cleanup: -ath11k_dp_srng_common_cleanup (sc); - fail_link_desc_cleanup:ath11k_dp_link_desc_cleanup (sc, dp->link_desc_banks,HAL_WBM_IDLE_LINK , &dp->wbm_idle_ring);
drivers/net/wireless/ath/ath11k/dp.h +0 -5
@@ -66,11 +66,6 @@ struct dp_tx_ring {u32 num_tx_pending;/* Protects txbuf_idr and num_pending */ spinlock_t tx_idr_lock; - DECLARE_KFIFO_PTR(tx_status_fifo, struct hal_wbm_release_ring); -/* lock to protect tx_status_fifo because tx_status_fifo can be -* accessed concurrently. -*/ -spinlock_t tx_status_lock; };struct ath11k_pdev_mon_stats {
drivers/net/wireless/ath/ath11k/dp_tx.c +8 -29
@@ -448,47 +448,28 @@ void ath11k_dp_tx_completion_handler(struct ath11k_base *ab, int ring_id)int hal_ring_id = dp->tx_ring[ring_id].tcl_comp_ring.ring_id;struct hal_srng *status_ring = &ab->hal.srng_list[hal_ring_id];struct sk_buff *msdu; -struct hal_wbm_release_ring tx_status;struct hal_tx_status ts;struct dp_tx_ring *tx_ring = &dp->tx_ring[ring_id]; -u32 *desc;u32 msdu_id; +u32 *desc;u8 mac_id;spin_lock_bh (&status_ring->lock);ath11k_hal_srng_access_begin (ab, status_ring); -spin_lock_bh (&tx_ring->tx_status_lock); -while (!kfifo_is_full (&tx_ring->tx_status_fifo) && - (desc =ath11k_hal_srng_dst_get_next_entry (ab, status_ring))) { -ath11k_hal_tx_status_desc_sync ((void *)desc, - (void *)&tx_status); -kfifo_put (&tx_ring->tx_status_fifo, tx_status); - } - -if ((ath11k_hal_srng_dst_peek (ab, status_ring) !=NULL ) && -kfifo_is_full (&tx_ring->tx_status_fifo)) { -/* TODO: Process pending tx_status messages when kfifo_is_full() */ -ath11k_warn (ab,"Unable to process some of the tx_status ring desc because status_fifo is full\n" ); - } - -spin_unlock_bh (&tx_ring->tx_status_lock); - -ath11k_hal_srng_access_end (ab, status_ring); -spin_unlock_bh (&status_ring->lock); - -spin_lock_bh (&tx_ring->tx_status_lock); -while (kfifo_get (&tx_ring->tx_status_fifo, &tx_status)) { +while ((desc =ath11k_hal_srng_dst_get_next_entry (ab, status_ring))) {memset (&ts,0 ,sizeof (ts)); -ath11k_hal_tx_status_parse (ab, &tx_status, &ts); +ath11k_hal_tx_status_parse (ab, + (struct hal_wbm_release_ring *)desc, + &ts); mac_id =FIELD_GET (DP_TX_DESC_ID_MAC_ID , ts.desc_id); msdu_id =FIELD_GET (DP_TX_DESC_ID_MSDU_ID , ts.desc_id);if (ts.buf_rel_source ==HAL_WBM_REL_SRC_MODULE_FW ) {ath11k_dp_tx_process_htt_tx_complete (ab, - (void *)&tx_status, + (void *)desc, mac_id, msdu_id, tx_ring);continue ;
@@ -511,12 +492,10 @@ void ath11k_dp_tx_completion_handler(struct ath11k_base *ab, int ring_id)if (atomic_dec_and_test (&ar->dp.num_tx_pending))wake_up (&ar->dp.tx_empty_waitq); -/* TODO: Locking optimization so that tx_completion for an msdu -* is not called with tx_status_lock acquired -*/ ath11k_dp_tx_complete_msdu (ar, msdu, &ts); } -spin_unlock_bh (&tx_ring->tx_status_lock); +ath11k_hal_srng_access_end (ab, status_ring); +spin_unlock_bh (&status_ring->lock); } int ath11k_dp_tx_send_reo_cmd(struct ath11k_base *ab, struct dp_rx_tid *rx_tid,
drivers/net/wireless/ath/ath11k/hal_tx.c +0 -5
@@ -79,11 +79,6 @@ void ath11k_hal_tx_desc_sync(void *tx_desc_cached, void *hw_desc) sizeof(struct hal_tcl_data_cmd)); } -void ath11k_hal_tx_status_desc_sync (void *hw_desc,void *local_desc) -{ -memcpy (local_desc, hw_desc,HAL_TX_STATUS_DESC_LEN ); -} -void ath11k_hal_tx_status_parse (struct ath11k_base *ab,struct hal_wbm_release_ring *desc,struct hal_tx_status *ts)
drivers/net/wireless/ath/ath11k/hal_tx.h +0 -1
@@ -70,7 +70,6 @@ void ath11k_hal_tx_desc_sync(void *tx_desc_cached, void *hw_desc);void ath11k_hal_tx_status_parse (struct ath11k_base *ab,struct hal_wbm_release_ring *desc,struct hal_tx_status *ts); -void ath11k_hal_tx_status_desc_sync (void *hw_desc,void *local_desc);void ath11k_hal_tx_set_dscp_tid_map (struct ath11k_base *ab,int id); int ath11k_hal_reo_cmd_send(struct ath11k_base *ab, struct hal_srng *srng, enum hal_reo_cmd_type type,
Signature · 7 lines
2.20.1 _______________________________________________ ath11k mailing list [email protected] http://lists.infradead.org/mailman/listinfo/ath11k
RFC 2/2 ath11k: drop memset in tx status hotpath
This in a small but measurable performance increase.
--- drivers/net/wireless/ath/ath11k/dp_tx.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/net/wireless/ath/ath11k/dp_tx.c +2 -2
@@ -448,7 +448,6 @@ void ath11k_dp_tx_completion_handler(struct ath11k_base *ab, int ring_id)int hal_ring_id = dp->tx_ring[ring_id].tcl_comp_ring.ring_id;struct hal_srng *status_ring = &ab->hal.srng_list[hal_ring_id];struct sk_buff *msdu; -struct hal_tx_status ts;struct dp_tx_ring *tx_ring = &dp->tx_ring[ring_id];u32 msdu_id;u32 *desc;
@@ -459,7 +458,8 @@ void ath11k_dp_tx_completion_handler(struct ath11k_base *ab, int ring_id)ath11k_hal_srng_access_begin (ab, status_ring);while ((desc =ath11k_hal_srng_dst_get_next_entry (ab, status_ring))) { -memset (&ts,0 ,sizeof (ts)); +struct hal_tx_status ts = {0 }; +ath11k_hal_tx_status_parse (ab, (struct hal_wbm_release_ring *)desc, &ts);
Signature · 7 lines
2.20.1 _______________________________________________ ath11k mailing list [email protected] http://lists.infradead.org/mailman/listinfo/ath11k
Re: [RFC 2/2] ath11k: drop memset in tx status hotpath
Quoted 20 lines · John Crispin
This in a small but measurable performance increase.
Signed-off-by John Crispin <[email protected]>--- drivers/net/wireless/ath/ath11k/dp_tx.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)drivers/net/wireless/ath/ath11k/dp_tx.c +0
-1@@ -448,7 +448,6 @@ void ath11k_dp_tx_completion_handler(struct ath11k_base *ab, int ring_id)int hal_ring_id = dp->tx_ring[ring_id].tcl_comp_ring.ring_id;struct hal_srng *status_ring = &ab->hal.srng_list[hal_ring_id];struct sk_buff *msdu; -struct hal_tx_status ts;struct dp_tx_ring *tx_ring = &dp->tx_ring[ring_id];u32 msdu_id;u32 *desc;
drivers/net/wireless/ath/ath11k/dp_tx.c void ath11k_dp_tx_completion_handler(struct ath11k_base *ab, int ring_id) · line 461
Quoted 5 lines
drivers/net/wireless/ath/ath11k/dp_tx.c +1
-1@@ -459,7 +458,8 @@ void ath11k_dp_tx_completion_handler(struct ath11k_base *ab, int ring_id)ath11k_hal_srng_access_begin (ab, status_ring);while ((desc =ath11k_hal_srng_dst_get_next_entry (ab, status_ring))) { -memset (&ts,0 ,sizeof (ts)); +struct hal_tx_status ts = {0 };
Very interesting, so a compiler zeroes memory from stack faster than memset() from "heap" (or whatever kmalloced memory should be called)? Just out of curiosity, any ideas why memset() is so slow?
Also some kind of numbers (before and after) would be nice to have in the commit log. It's not uncommon that years after we investigate why a certain change was made.
Signature · 6 lines
Kalle Valo _______________________________________________ ath11k mailing list [email protected] http://lists.infradead.org/mailman/listinfo/ath11k
Quoted 121 lines · John Crispin
perf top showed that 10% of the cpu time was spent on the memcpy used to populate the fifo when the device is under high load. There is no apparent benefit from using a fifo so lets just drop it.
Signed-off-by John Crispin <[email protected]>--- drivers/net/wireless/ath/ath11k/dp.c | 13 --------- drivers/net/wireless/ath/ath11k/dp.h | 5 ---- drivers/net/wireless/ath/ath11k/dp_tx.c | 37 +++++------------------- drivers/net/wireless/ath/ath11k/hal_tx.c | 5 ---- drivers/net/wireless/ath/ath11k/hal_tx.h | 1 - 5 files changed, 8 insertions(+), 53 deletions(-)drivers/net/wireless/ath/ath11k/dp.c +0
-13@@ -839,10 +839,6 @@ void ath11k_dp_free(struct ath11k_base *sc) ath11k_dp_tx_pending_cleanup, sc);idr_destroy (&dp->tx_ring[i].txbuf_idr);spin_unlock_bh (&dp->tx_ring[i].tx_idr_lock); - -spin_lock_bh (&dp->tx_ring[i].tx_status_lock); -kfifo_free (&dp->tx_ring[i].tx_status_fifo); -spin_unlock_bh (&dp->tx_ring[i].tx_status_lock); }/* Deinit any SOC level resource */ @@ -888,12 +884,6 @@ int ath11k_dp_alloc(struct ath11k_base *sc)idr_init (&dp->tx_ring[i].txbuf_idr);spin_lock_init (&dp->tx_ring[i].tx_idr_lock); dp->tx_ring[i].tcl_data_ring_id = i; - -spin_lock_init (&dp->tx_ring[i].tx_status_lock); - ret =kfifo_alloc (&dp->tx_ring[i].tx_status_fifo, size, -GFP_KERNEL ); -if (ret) - goto fail_cmn_srng_cleanup; }for (i =0 ; i <HAL_DSCP_TID_MAP_TBL_NUM_ENTRIES_MAX ; i++)@@ -903,9 +893,6 @@ int ath11k_dp_alloc(struct ath11k_base *sc)return 0 ; -fail_cmn_srng_cleanup: -ath11k_dp_srng_common_cleanup (sc); - fail_link_desc_cleanup:ath11k_dp_link_desc_cleanup (sc, dp->link_desc_banks,HAL_WBM_IDLE_LINK , &dp->wbm_idle_ring);drivers/net/wireless/ath/ath11k/dp.h +0
-5@@ -66,11 +66,6 @@ struct dp_tx_ring {u32 num_tx_pending;/* Protects txbuf_idr and num_pending */ spinlock_t tx_idr_lock; - DECLARE_KFIFO_PTR(tx_status_fifo, struct hal_wbm_release_ring); -/* lock to protect tx_status_fifo because tx_status_fifo can be -* accessed concurrently. -*/ -spinlock_t tx_status_lock; };struct ath11k_pdev_mon_stats {drivers/net/wireless/ath/ath11k/dp_tx.c +5
-24@@ -448,47 +448,28 @@ void ath11k_dp_tx_completion_handler(struct ath11k_base *ab, int ring_id) int hal_ring_id = dp->tx_ring[ring_id].tcl_comp_ring.ring_id;struct hal_srng *status_ring = &ab->hal.srng_list[hal_ring_id];struct sk_buff *msdu; -struct hal_wbm_release_ring tx_status;struct hal_tx_status ts;struct dp_tx_ring *tx_ring = &dp->tx_ring[ring_id]; -u32 *desc;u32 msdu_id; +u32 *desc;u8 mac_id;spin_lock_bh (&status_ring->lock);ath11k_hal_srng_access_begin (ab, status_ring); -spin_lock_bh (&tx_ring->tx_status_lock); -while (!kfifo_is_full (&tx_ring->tx_status_fifo) && - (desc =ath11k_hal_srng_dst_get_next_entry (ab, status_ring))) { -ath11k_hal_tx_status_desc_sync ((void *)desc, - (void *)&tx_status); -kfifo_put (&tx_ring->tx_status_fifo, tx_status); - } - -if ((ath11k_hal_srng_dst_peek (ab, status_ring) !=NULL ) && -kfifo_is_full (&tx_ring->tx_status_fifo)) { -/* TODO: Process pending tx_status messages when kfifo_is_full() */ - ath11k_warn(ab, "Unable to process some of the tx_status ring desc because status_fifo is full\n"); - } - - spin_unlock_bh(&tx_ring->tx_status_lock); - -ath11k_hal_srng_access_end (ab, status_ring); -spin_unlock_bh (&status_ring->lock); - -spin_lock_bh (&tx_ring->tx_status_lock); -while (kfifo_get (&tx_ring->tx_status_fifo, &tx_status)) { +while ((desc =ath11k_hal_srng_dst_get_next_entry (ab, status_ring))) {memset (&ts,0 ,sizeof (ts)); -ath11k_hal_tx_status_parse (ab, &tx_status, &ts); +ath11k_hal_tx_status_parse (ab, + (struct hal_wbm_release_ring *)desc, + &ts);
drivers/net/wireless/ath/ath11k/dp_tx.c void ath11k_dp_tx_completion_handler(struct · line 474
Quoted 10 lines
drivers/net/wireless/ath/ath11k/dp_tx.c +1
-1mac_id =FIELD_GET (DP_TX_DESC_ID_MAC_ID , ts.desc_id); msdu_id =FIELD_GET (DP_TX_DESC_ID_MSDU_ID , ts.desc_id);if (ts.buf_rel_source ==HAL_WBM_REL_SRC_MODULE_FW ) {ath11k_dp_tx_process_htt_tx_complete (ab, - (void *)&tx_status, + (void *)desc, mac_id, msdu_id, tx_ring);
Delay in reaping tx completion desc from hw ring could trigger backpressure on tx completion ring which could result in target assert. Extensive testing with this change is required to make sure this change does not cause any issue.
Vasanth
_______________________________________________
ath11k mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/ath11k