Skip to contentSkip to conversation

rsi: Move card interrupt handling to RX thread

PATCH In review Marek Vasut 3 participants 5 messages Nov 3, 2020 – Nov 10, 2020

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/2 rsi: Move card interrupt handling to RX thread

The interrupt handling of the RS911x is particularly heavy. For each RX packet, the card does three SDIO transactions, one to read interrupt status register, one to RX buffer length, one to read the RX packet(s). This translates to ~330 uS per one cycle of interrupt handler. In case there is more incoming traffic, this will be more.

The drivers/mmc/core/sdio_irq.c has the following comment, quote "Just like traditional hard IRQ handlers, we expect SDIO IRQ handlers to be quick and to the point, so that the holding of the host lock does not cover too much work that doesn't require that lock to be held."

The RS911x interrupt handler does not fit that. This patch therefore changes it such that the entire IRQ handler is moved to the RX thread instead, and the interrupt handler only wakes the RX thread.

This is OK, because the interrupt handler only does things which can also be done in the RX thread, that is, it checks for firmware loading error(s), it checks buffer status, it checks whether a packet arrived and if so, reads out the packet and passes it to network stack.

Moreover, this change permits removal of a code which allocated an skbuff only to get 4-byte-aligned buffer, read up to 8kiB of data into the skbuff, queue this skbuff into local private queue, then in RX thread, this buffer is dequeued, the data in the skbuff as passed to the RSI driver core, and the skbuff is deallocated. All this is replaced by directly calling the RSI driver core with local buffer.

Cc Angus Ainslie <[email protected]>
Cc David S. Miller <[email protected]>
Cc Jakub Kicinski <[email protected]>
Cc Kalle Valo <[email protected]>
Cc Lee Jones <[email protected]>
Cc Martin Kepplinger <[email protected]>
Cc Sebastian Krzyszkowiak <[email protected]>
Cc Siva Rebbagondla <[email protected]>
---
 drivers/net/wireless/rsi/rsi_91x_sdio.c     |  6 +--
 drivers/net/wireless/rsi/rsi_91x_sdio_ops.c | 52 ++++++---------------
 drivers/net/wireless/rsi/rsi_sdio.h         |  8 +---
 3 files changed, 15 insertions(+), 51 deletions(-)
drivers/net/wireless/rsi/rsi_91x_sdio.c +1 -5
index 78ad605081ae..d81874eee4b3 100644
@@ -153,9 +153,7 @@ static void rsi_handle_interrupt(struct sdio_func *function)
 	if (adapter->priv->fsm_state == FSM_FW_NOT_LOADED)
 		return;
 
-	dev->sdio_irq_task = current;
-	rsi_interrupt_handler(adapter);
-	dev->sdio_irq_task = NULL;
+	rsi_set_event(&dev->rx_thread.event);
 }
 
 /**
@@ -1058,8 +1056,6 @@ static int rsi_probe(struct sdio_func *pfunction,
 		rsi_dbg(ERR_ZONE, "%s: Unable to init rx thrd\n", __func__);
 		goto fail_kill_thread;
 	}
-	skb_queue_head_init(&sdev->rx_q.head);
-	sdev->rx_q.num_rx_pkts = 0;
 
 	sdio_claim_host(pfunction);
 	if (sdio_claim_irq(pfunction, rsi_handle_interrupt)) {
drivers/net/wireless/rsi/rsi_91x_sdio_ops.c +13 -39

96 lines not on this page · raw message

drivers/net/wireless/rsi/rsi_sdio.h +1 -7
index 9afc1d0d2684..1c756263cf15 100644
@@ -107,11 +107,6 @@ struct receive_info {
 	u32 buf_available_counter;
 };
 
-struct rsi_sdio_rx_q {
-	u8 num_rx_pkts;
-	struct sk_buff_head head;
-};
-
 struct rsi_91x_sdiodev {
 	struct sdio_func *pfunction;
 	struct task_struct *sdio_irq_task;
@@ -124,11 +119,10 @@ struct rsi_91x_sdiodev {
 	u16 tx_blk_size;
 	u8 write_fail;
 	bool buff_status_updated;
-	struct rsi_sdio_rx_q rx_q;
 	struct rsi_thread rx_thread;
+	u8 pktbuffer[8192] __aligned(4);
 };
 
-void rsi_interrupt_handler(struct rsi_hw *adapter);
 int rsi_init_sdio_slave_regs(struct rsi_hw *adapter);
 int rsi_sdio_read_register(struct rsi_hw *adapter, u32 addr, u8 *data);
 int rsi_sdio_host_intf_read_pkt(struct rsi_hw *adapter, u8 *pkt, u32 length);
Signature · 1 line
2.28.0
#1 lore raw

PATCH 2/2 rsi: Clean up loop in the interrupt handler

The inner do { ... } while loop is completely useless, all it does is iterate over a switch-case statement, one bit at a time. This can easily be replaced by simple if (status & bit) { ... } tests for each bit. No functional change.

Cc Angus Ainslie <[email protected]>
Cc David S. Miller <[email protected]>
Cc Jakub Kicinski <[email protected]>
Cc Kalle Valo <[email protected]>
Cc Lee Jones <[email protected]>
Cc Martin Kepplinger <[email protected]>
Cc Sebastian Krzyszkowiak <[email protected]>
Cc Siva Rebbagondla <[email protected]>
---
 drivers/net/wireless/rsi/rsi_91x_sdio_ops.c | 121 ++++++++++----------
 1 file changed, 58 insertions(+), 63 deletions(-)
drivers/net/wireless/rsi/rsi_91x_sdio_ops.c +58 -63
index 23e709aabd1f..8ace1874e5cb 100644
@@ -236,7 +236,6 @@ static void rsi_rx_handler(struct rsi_hw *adapter)
 	struct rsi_91x_sdiodev *dev =
 		(struct rsi_91x_sdiodev *)adapter->rsi_dev;
 	int status;
-	enum sdio_interrupt_type isr_type;
 	u8 isr_status = 0;
 	u8 fw_status = 0;
 
@@ -267,73 +266,69 @@ static void rsi_rx_handler(struct rsi_hw *adapter)
 			__func__, isr_status, (1 << MSDU_PKT_PENDING),
 			(1 << FW_ASSERT_IND));
 
-		do {
-			RSI_GET_SDIO_INTERRUPT_TYPE(isr_status, isr_type);
-
-			switch (isr_type) {
-			case BUFFER_AVAILABLE:
-				status = rsi_sdio_check_buffer_status(adapter,
-								      0);
-				if (status < 0)
-					rsi_dbg(ERR_ZONE,
-						"%s: Failed to check buffer status\n",
-						__func__);
-				rsi_sdio_ack_intr(common->priv,
-						  (1 << PKT_BUFF_AVAILABLE));
-				rsi_set_event(&common->tx_thread.event);
-
-				rsi_dbg(ISR_ZONE,
-					"%s: ==> BUFFER_AVAILABLE <==\n",
-					__func__);
-				dev->buff_status_updated = true;
-				break;
-
-			case FIRMWARE_ASSERT_IND:
+		if (isr_status & BIT(PKT_BUFF_AVAILABLE)) {
+			status = rsi_sdio_check_buffer_status(adapter, 0);
+			if (status < 0)
 				rsi_dbg(ERR_ZONE,
-					"%s: ==> FIRMWARE Assert <==\n",
+					"%s: Failed to check buffer status\n",
 					__func__);
-				status = rsi_sdio_read_register(common->priv,
+			rsi_sdio_ack_intr(common->priv,
+					  BIT(PKT_BUFF_AVAILABLE));
+			rsi_set_event(&common->tx_thread.event);
+
+			rsi_dbg(ISR_ZONE, "%s: ==> BUFFER_AVAILABLE <==\n",
+				__func__);
+			dev->buff_status_updated = true;
+
+			isr_status &= ~BIT(PKT_BUFF_AVAILABLE);
+		}
+
+		if (isr_status & BIT(FW_ASSERT_IND)) {
+			rsi_dbg(ERR_ZONE, "%s: ==> FIRMWARE Assert <==\n",
+				__func__);
+			status = rsi_sdio_read_register(common->priv,
 							SDIO_FW_STATUS_REG,
 							&fw_status);
-				if (status) {
-					rsi_dbg(ERR_ZONE,
-						"%s: Failed to read f/w reg\n",
-						__func__);
-				} else {
-					rsi_dbg(ERR_ZONE,
-						"%s: Firmware Status is 0x%x\n",
-						__func__ , fw_status);
-					rsi_sdio_ack_intr(common->priv,
-							  (1 << FW_ASSERT_IND));
-				}
-
-				common->fsm_state = FSM_CARD_NOT_READY;
-				break;
-
-			case MSDU_PACKET_PENDING:
-				rsi_dbg(ISR_ZONE, "Pkt pending interrupt\n");
-				dev->rx_info.total_sdio_msdu_pending_intr++;
-
-				status = rsi_process_pkt(common);
-				if (status) {
-					rsi_dbg(ERR_ZONE,
-						"%s: Failed to read pkt\n",
-						__func__);
-					mutex_unlock(&common->rx_lock);
-					return;
-				}
-				break;
-			default:
-				rsi_sdio_ack_intr(common->priv, isr_status);
-				dev->rx_info.total_sdio_unknown_intr++;
-				isr_status = 0;
-				rsi_dbg(ISR_ZONE,
-					"Unknown Interrupt %x\n",
-					isr_status);
-				break;
+			if (status) {
+				rsi_dbg(ERR_ZONE,
+					"%s: Failed to read f/w reg\n",
+					__func__);
+			} else {
+				rsi_dbg(ERR_ZONE,
+					"%s: Firmware Status is 0x%x\n",
+					__func__, fw_status);
+				rsi_sdio_ack_intr(common->priv,
+						  BIT(FW_ASSERT_IND));
 			}
-			isr_status ^= BIT(isr_type - 1);
-		} while (isr_status);
+
+			common->fsm_state = FSM_CARD_NOT_READY;
+
+			isr_status &= ~BIT(FW_ASSERT_IND);
+		}
+
+		if (isr_status & BIT(MSDU_PKT_PENDING)) {
+			rsi_dbg(ISR_ZONE, "Pkt pending interrupt\n");
+			dev->rx_info.total_sdio_msdu_pending_intr++;
+
+			status = rsi_process_pkt(common);
+			if (status) {
+				rsi_dbg(ERR_ZONE, "%s: Failed to read pkt\n",
+					__func__);
+				mutex_unlock(&common->rx_lock);
+				return;
+			}
+
+			isr_status &= ~BIT(MSDU_PKT_PENDING);
+		}
+
+		if (isr_status) {
+			rsi_sdio_ack_intr(common->priv, isr_status);
+			dev->rx_info.total_sdio_unknown_intr++;
+			isr_status = 0;
+			rsi_dbg(ISR_ZONE, "Unknown Interrupt %x\n",
+				isr_status);
+		}
+
 		mutex_unlock(&common->rx_lock);
 	} while (1);
 }
Signature · 1 line
2.28.0
Tested-by on v1 1/2
#2 lore raw
In reply to Marek Vasut
Quoted 38 lines · Marek Vasut

The interrupt handling of the RS911x is particularly heavy. For each RX packet, the card does three SDIO transactions, one to read interrupt status register, one to RX buffer length, one to read the RX packet(s). This translates to ~330 uS per one cycle of interrupt handler. In case there is more incoming traffic, this will be more.

The drivers/mmc/core/sdio_irq.c has the following comment, quote "Just like traditional hard IRQ handlers, we expect SDIO IRQ handlers to be quick and to the point, so that the holding of the host lock does not cover too much work that doesn't require that lock to be held."

The RS911x interrupt handler does not fit that. This patch therefore changes it such that the entire IRQ handler is moved to the RX thread instead, and the interrupt handler only wakes the RX thread.

This is OK, because the interrupt handler only does things which can also be done in the RX thread, that is, it checks for firmware loading error(s), it checks buffer status, it checks whether a packet arrived and if so, reads out the packet and passes it to network stack.

Moreover, this change permits removal of a code which allocated an skbuff only to get 4-byte-aligned buffer, read up to 8kiB of data into the skbuff, queue this skbuff into local private queue, then in RX thread, this buffer is dequeued, the data in the skbuff as passed to the RSI driver core, and the skbuff is deallocated. All this is replaced by directly calling the RSI driver core with local buffer.

Cc Angus Ainslie <[email protected]>
Cc David S. Miller <[email protected]>
Cc Jakub Kicinski <[email protected]>
Cc Kalle Valo <[email protected]>
Cc Lee Jones <[email protected]>
Cc Martin Kepplinger <[email protected]>
Cc Sebastian Krzyszkowiak <[email protected]>
Cc Siva Rebbagondla <[email protected]>
Quoted 5 lines
---
  drivers/net/wireless/rsi/rsi_91x_sdio.c     |  6 +--
  drivers/net/wireless/rsi/rsi_91x_sdio_ops.c | 52 ++++++---------------
  drivers/net/wireless/rsi/rsi_sdio.h         |  8 +---
  3 files changed, 15 insertions(+), 51 deletions(-)

hi Marek,

I'm running this without problems, so feel free to add

to both of your changes. thanks a lot,

martin

#3 lore raw
Quoted 46 lines · Martin Kepplinger
Quoted 38 lines · Marek Vasut

The interrupt handling of the RS911x is particularly heavy. For each RX packet, the card does three SDIO transactions, one to read interrupt status register, one to RX buffer length, one to read the RX packet(s). This translates to ~330 uS per one cycle of interrupt handler. In case there is more incoming traffic, this will be more.

The drivers/mmc/core/sdio_irq.c has the following comment, quote "Just like traditional hard IRQ handlers, we expect SDIO IRQ handlers to be quick and to the point, so that the holding of the host lock does not cover too much work that doesn't require that lock to be held."

The RS911x interrupt handler does not fit that. This patch therefore changes it such that the entire IRQ handler is moved to the RX thread instead, and the interrupt handler only wakes the RX thread.

This is OK, because the interrupt handler only does things which can also be done in the RX thread, that is, it checks for firmware loading error(s), it checks buffer status, it checks whether a packet arrived and if so, reads out the packet and passes it to network stack.

Moreover, this change permits removal of a code which allocated an skbuff only to get 4-byte-aligned buffer, read up to 8kiB of data into the skbuff, queue this skbuff into local private queue, then in RX thread, this buffer is dequeued, the data in the skbuff as passed to the RSI driver core, and the skbuff is deallocated. All this is replaced by directly calling the RSI driver core with local buffer.

Cc Angus Ainslie <[email protected]>
Cc David S. Miller <[email protected]>
Cc Jakub Kicinski <[email protected]>
Cc Kalle Valo <[email protected]>
Cc Lee Jones <[email protected]>
Cc Martin Kepplinger <[email protected]>
Cc Sebastian Krzyszkowiak <[email protected]>
Cc Siva Rebbagondla <[email protected]>
Quoted 5 lines

---
  drivers/net/wireless/rsi/rsi_91x_sdio.c     |  6 +-- drivers/net/wireless/rsi/rsi_91x_sdio_ops.c | 52 ++++++--------------- drivers/net/wireless/rsi/rsi_sdio.h         |  8 +---

  3 files changed, 15 insertions(+), 51 deletions(-)
Quoted 1 line

hi Marek,

Hi,

Quoted 3 lines

I'm running this without problems, so feel free to add

Thank you.

Do you observe better RX performance ? For me, without this patch, iperf3 -R did ~4 Mbit/s on iwlwifi AP running hostap in 802.11n mode (wpa2 tkip), with this patch I see 40 Mbit/s (10x better, yes). However, the poor rx performance did depend on the kernel configuration (HZ, preemption settings) before, now it does not.

#4 lore raw
In reply to Marek Vasut
Quoted 27 lines · Marek Vasut

The interrupt handling of the RS911x is particularly heavy. For each RX packet, the card does three SDIO transactions, one to read interrupt status register, one to RX buffer length, one to read the RX packet(s). This translates to ~330 uS per one cycle of interrupt handler. In case there is more incoming traffic, this will be more.

The drivers/mmc/core/sdio_irq.c has the following comment, quote "Just like traditional hard IRQ handlers, we expect SDIO IRQ handlers to be quick and to the point, so that the holding of the host lock does not cover too much work that doesn't require that lock to be held."

The RS911x interrupt handler does not fit that. This patch therefore changes it such that the entire IRQ handler is moved to the RX thread instead, and the interrupt handler only wakes the RX thread.

This is OK, because the interrupt handler only does things which can also be done in the RX thread, that is, it checks for firmware loading error(s), it checks buffer status, it checks whether a packet arrived and if so, reads out the packet and passes it to network stack.

Moreover, this change permits removal of a code which allocated an skbuff only to get 4-byte-aligned buffer, read up to 8kiB of data into the skbuff, queue this skbuff into local private queue, then in RX thread, this buffer is dequeued, the data in the skbuff as passed to the RSI driver core, and the skbuff is deallocated. All this is replaced by directly calling the RSI driver core with local buffer.

Quoted 12 lines
Cc Angus Ainslie <[email protected]>
Cc David S. Miller <[email protected]>
Cc Jakub Kicinski <[email protected]>
Cc Kalle Valo <[email protected]>
Cc Lee Jones <[email protected]>
Cc Martin Kepplinger <[email protected]>
Cc Sebastian Krzyszkowiak <[email protected]>
Cc Siva Rebbagondla <[email protected]>

2 patches applied to wireless-drivers-next.git, thanks.

287431463e78 rsi: Move card interrupt handling to RX thread abd131a19f6b rsi: Clean up loop in the interrupt handler

Signature · 3 lines
https://patchwork.kernel.org/project/linux-wireless/patch/[email protected]/

https://wireless.wiki.kernel.org/en/developers/documentation/submittingpatches