[PATCH V2 net-next 0/7] ENA driver new features

[PATCH V2 net-next 0/7] ENA driver new features

[akiyano]

From: Arthur Kiyanovski <akiyano@amazon.com>


V2 changes:
-----------
1. Update commit messages of 2 patches to be more verbose.
2. Remove "net: ena: enable support of rss hash key and function
   changes" patch. Will be resubmitted net.


V1 cover letter:
----------------
This patchset contains performance improvements, support for new devices
and functionality:

1. Support for upcoming ENA devices
2. Avoid unnecessary IRQ unmasking in busy poll to reduce interrupt rate
3. Enabling device support for RSS function and key manipulation
4. Support for NIC-based traffic mirroring (SPAN port)
5. Additional PCI device ID 
6. Cosmetic changes

Arthur Kiyanovski (7):
  net: ena: avoid unnecessary rearming of interrupt vector when
    busy-polling
  net: ena: add reserved PCI device ID
  net: ena: cosmetic: satisfy gcc warning
  net: ena: cosmetic: change ena_com_stats_admin stats to u64
  net: ena: add support for traffic mirroring
  net: ena: move llq configuration from ena_probe to ena_device_init()
  net: ena: support new LLQ acceleration mode

 .../net/ethernet/amazon/ena/ena_admin_defs.h  |  44 ++++-
 drivers/net/ethernet/amazon/ena/ena_com.c     |  19 +-
 drivers/net/ethernet/amazon/ena/ena_com.h     |  13 +-
 drivers/net/ethernet/amazon/ena/ena_eth_com.c |  51 +++--
 drivers/net/ethernet/amazon/ena/ena_eth_com.h |   3 +-
 drivers/net/ethernet/amazon/ena/ena_ethtool.c |   4 +-
 drivers/net/ethernet/amazon/ena/ena_netdev.c  | 176 ++++++++++--------
 drivers/net/ethernet/amazon/ena/ena_netdev.h  |   3 +
 .../net/ethernet/amazon/ena/ena_pci_id_tbl.h  |   5 +
 9 files changed, 215 insertions(+), 103 deletions(-)

-- 
2.23.1

[PATCH V2 net-next 1/7] net: ena: avoid unnecessary rearming of interrupt vector when busy-polling

[akiyano]

From: Arthur Kiyanovski <akiyano@amazon.com>

In napi busy-poll mode, the kernel invokes the napi handler of the
device repeatedly to poll the NIC's receive queues. This process
repeats until a timeout, specific for each connection, is up.
By polling packets in busy-poll mode the user may gain lower latency
and higher throughput (since the kernel no longer waits for interrupts
to poll the queues) in expense of CPU usage.

Upon completing a napi routine, the driver checks whether
the routine was called by an interrupt handler. If so, the driver
re-enables interrupts for the device. This is needed since an
interrupt routine invocation disables future invocations until
explicitly re-enabled.

The driver avoids re-enabling the interrupts if they were not disabled
in the first place (e.g. if driver in busy mode).
Originally, the driver checked whether interrupt re-enabling is needed
by reading the 'ena_napi->unmask_interrupt' variable. This atomic
variable was set upon interrupt and cleared after re-enabling it.

In the 4.10 Linux version, the 'napi_complete_done' call was changed
so that it returns 'false' when device should not re-enable
interrupts, and 'true' otherwise. The change includes reading the
"NAPIF_STATE_IN_BUSY_POLL" flag to check if the napi call is in
busy-poll mode, and if so, return 'false'.
The driver was changed to re-enable interrupts according to this
routine's return value.
The Linux community rejected the use of the
'ena_napi->unmaunmask_interrupt' variable to determine whether
unmasking is needed, and urged to use napi_napi_complete_done()
return value solely.
See https://lore.kernel.org/patchwork/patch/741149/ for more details

As explained, a busy-poll session exists for a specified timeout
value, after which it exits the busy-poll mode and re-enters it later.
This leads to many invocations of the napi handler where
napi_complete_done() false indicates that interrupts should be
re-enabled.
This creates a bug in which the interrupts are re-enabled
unnecessarily.
To reproduce this bug:
    1) echo 50 | sudo tee /proc/sys/net/core/busy_poll
    2) echo 50 | sudo tee /proc/sys/net/core/busy_read
    3) Add counters that check whether
    'ena_unmask_interrupt(tx_ring, rx_ring);'
    is called without disabling the interrupts in the first
    place (i.e. with calling the interrupt routine
    ena_intr_msix_io())

Steps 1+2 enable busy-poll as the default mode for new connections.

The busy poll routine rearms the interrupts after every session by
design, and so we need to add an extra check that the interrupts were
masked in the first place.

Signed-off-by: Shay Agroskin <shayagr@amazon.com>
Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
---
 drivers/net/ethernet/amazon/ena/ena_netdev.c | 7 ++++++-
 drivers/net/ethernet/amazon/ena/ena_netdev.h | 1 +
 2 files changed, 7 insertions(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/amazon/ena/ena_netdev.c b/drivers/net/ethernet/amazon/ena/ena_netdev.c
index 91be3ffa1c5c..90c0fe15cd23 100644
--- a/drivers/net/ethernet/amazon/ena/ena_netdev.c
+++ b/drivers/net/ethernet/amazon/ena/ena_netdev.c
@@ -1913,7 +1913,9 @@ static int ena_io_poll(struct napi_struct *napi, int budget)
 		/* Update numa and unmask the interrupt only when schedule
 		 * from the interrupt context (vs from sk_busy_loop)
 		 */
-		if (napi_complete_done(napi, rx_work_done)) {
+		if (napi_complete_done(napi, rx_work_done) &&
+		    READ_ONCE(ena_napi->interrupts_masked)) {
+			WRITE_ONCE(ena_napi->interrupts_masked, false);
 			/* We apply adaptive moderation on Rx path only.
 			 * Tx uses static interrupt moderation.
 			 */
@@ -1961,6 +1963,9 @@ static irqreturn_t ena_intr_msix_io(int irq, void *data)
 
 	ena_napi->first_interrupt = true;
 
+	WRITE_ONCE(ena_napi->interrupts_masked, true);
+	smp_wmb(); /* write interrupts_masked before calling napi */
+
 	napi_schedule_irqoff(&ena_napi->napi);
 
 	return IRQ_HANDLED;
diff --git a/drivers/net/ethernet/amazon/ena/ena_netdev.h b/drivers/net/ethernet/amazon/ena/ena_netdev.h
index ba030d260940..89304b403995 100644
--- a/drivers/net/ethernet/amazon/ena/ena_netdev.h
+++ b/drivers/net/ethernet/amazon/ena/ena_netdev.h
@@ -167,6 +167,7 @@ struct ena_napi {
 	struct ena_ring *rx_ring;
 	struct ena_ring *xdp_ring;
 	bool first_interrupt;
+	bool interrupts_masked;
 	u32 qid;
 	struct dim dim;
 };
-- 
2.23.1

[PATCH V2 net-next 2/7] net: ena: add reserved PCI device ID

[akiyano]

From: Arthur Kiyanovski <akiyano@amazon.com>

Add a reserved PCI device ID to the driver's table.
Used for internal testing purposes.

Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
---
 drivers/net/ethernet/amazon/ena/ena_pci_id_tbl.h | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/drivers/net/ethernet/amazon/ena/ena_pci_id_tbl.h b/drivers/net/ethernet/amazon/ena/ena_pci_id_tbl.h
index f80d2a47fa94..426e57e10a7f 100644
--- a/drivers/net/ethernet/amazon/ena/ena_pci_id_tbl.h
+++ b/drivers/net/ethernet/amazon/ena/ena_pci_id_tbl.h
@@ -53,10 +53,15 @@
 #define PCI_DEV_ID_ENA_LLQ_VF	0xec21
 #endif
 
+#ifndef PCI_DEV_ID_ENA_RESRV0
+#define PCI_DEV_ID_ENA_RESRV0	0x0051
+#endif
+
 #define ENA_PCI_ID_TABLE_ENTRY(devid) \
 	{PCI_DEVICE(PCI_VENDOR_ID_AMAZON, devid)},
 
 static const struct pci_device_id ena_pci_tbl[] = {
+	ENA_PCI_ID_TABLE_ENTRY(PCI_DEV_ID_ENA_RESRV0)
 	ENA_PCI_ID_TABLE_ENTRY(PCI_DEV_ID_ENA_PF)
 	ENA_PCI_ID_TABLE_ENTRY(PCI_DEV_ID_ENA_LLQ_PF)
 	ENA_PCI_ID_TABLE_ENTRY(PCI_DEV_ID_ENA_VF)
-- 
2.23.1

[PATCH V2 net-next 3/7] net: ena: cosmetic: satisfy gcc warning

[akiyano]

From: Arthur Kiyanovski <akiyano@amazon.com>

gcc 4.8 reports a warning when initializing with = {0}.
Dropping the "0" from the braces fixes the issue.
This fix is not ANSI compatible but is allowed by gcc.

Signed-off-by: Sameeh Jubran <sameehj@amazon.com>
Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
---
 drivers/net/ethernet/amazon/ena/ena_netdev.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/amazon/ena/ena_netdev.c b/drivers/net/ethernet/amazon/ena/ena_netdev.c
index 90c0fe15cd23..18a30a81a475 100644
--- a/drivers/net/ethernet/amazon/ena/ena_netdev.c
+++ b/drivers/net/ethernet/amazon/ena/ena_netdev.c
@@ -307,7 +307,7 @@ static int ena_xdp_xmit_buff(struct net_device *dev,
 			     struct ena_rx_buffer *rx_info)
 {
 	struct ena_adapter *adapter = netdev_priv(dev);
-	struct ena_com_tx_ctx ena_tx_ctx = {0};
+	struct ena_com_tx_ctx ena_tx_ctx = {};
 	struct ena_tx_buffer *tx_info;
 	struct ena_ring *xdp_ring;
 	u16 next_to_use, req_id;
-- 
2.23.1

[PATCH V2 net-next 4/7] net: ena: cosmetic: change ena_com_stats_admin stats to u64

[akiyano]

From: Arthur Kiyanovski <akiyano@amazon.com>

The size of the admin statistics in ena_com_stats_admin is changed
from 32bit to 64bit so to align with the sizes of the other statistics
in the driver (i.e. rx_stats, tx_stats and ena_stats_dev).

This is done as part of an effort to create a unified API to read
statistics.

Signed-off-by: Shay Agroskin <shayagr@amazon.com>
Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
---
 drivers/net/ethernet/amazon/ena/ena_com.h     | 10 +++++-----
 drivers/net/ethernet/amazon/ena/ena_ethtool.c |  4 ++--
 2 files changed, 7 insertions(+), 7 deletions(-)

diff --git a/drivers/net/ethernet/amazon/ena/ena_com.h b/drivers/net/ethernet/amazon/ena/ena_com.h
index bc187adf54e4..4c98f6f07882 100644
--- a/drivers/net/ethernet/amazon/ena/ena_com.h
+++ b/drivers/net/ethernet/amazon/ena/ena_com.h
@@ -230,11 +230,11 @@ struct ena_com_admin_sq {
 };
 
 struct ena_com_stats_admin {
-	u32 aborted_cmd;
-	u32 submitted_cmd;
-	u32 completed_cmd;
-	u32 out_of_space;
-	u32 no_completion;
+	u64 aborted_cmd;
+	u64 submitted_cmd;
+	u64 completed_cmd;
+	u64 out_of_space;
+	u64 no_completion;
 };
 
 struct ena_com_admin_queue {
diff --git a/drivers/net/ethernet/amazon/ena/ena_ethtool.c b/drivers/net/ethernet/amazon/ena/ena_ethtool.c
index e340b65af08c..430275bc0d04 100644
--- a/drivers/net/ethernet/amazon/ena/ena_ethtool.c
+++ b/drivers/net/ethernet/amazon/ena/ena_ethtool.c
@@ -164,13 +164,13 @@ static void ena_queue_stats(struct ena_adapter *adapter, u64 **data)
 static void ena_dev_admin_queue_stats(struct ena_adapter *adapter, u64 **data)
 {
 	const struct ena_stats *ena_stats;
-	u32 *ptr;
+	u64 *ptr;
 	int i;
 
 	for (i = 0; i < ENA_STATS_ARRAY_ENA_COM; i++) {
 		ena_stats = &ena_stats_ena_com_strings[i];
 
-		ptr = (u32 *)((uintptr_t)&adapter->ena_dev->admin_queue.stats +
+		ptr = (u64 *)((uintptr_t)&adapter->ena_dev->admin_queue.stats +
 			(uintptr_t)ena_stats->stat_offset);
 
 		*(*data)++ = *ptr;
-- 
2.23.1

[PATCH V2 net-next 5/7] net: ena: add support for traffic mirroring

[akiyano]

From: Arthur Kiyanovski <akiyano@amazon.com>

Add support for traffic mirroring, where the hardware reads the
buffer from the instance memory directly.

Traffic Mirroring needs access to the rx buffers in the instance.
To have this access, this patch:
1. Changes the code to map and unmap the rx buffers bidirectionally.
2. Enables the relevant bit in driver_supported_features to indicate
   to the FW that this driver supports traffic mirroring.

Rx completion is not generated until mirroring is done to avoid
the situation where the driver changes the buffer before it is
mirrored.

Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
---
 drivers/net/ethernet/amazon/ena/ena_admin_defs.h |  5 ++++-
 drivers/net/ethernet/amazon/ena/ena_netdev.c     | 15 +++++++++------
 2 files changed, 13 insertions(+), 7 deletions(-)

diff --git a/drivers/net/ethernet/amazon/ena/ena_admin_defs.h b/drivers/net/ethernet/amazon/ena/ena_admin_defs.h
index 336742f6e3c3..afe87ff09b20 100644
--- a/drivers/net/ethernet/amazon/ena/ena_admin_defs.h
+++ b/drivers/net/ethernet/amazon/ena/ena_admin_defs.h
@@ -816,7 +816,8 @@ struct ena_admin_host_info {
 	/* 0 : reserved
 	 * 1 : rx_offset
 	 * 2 : interrupt_moderation
-	 * 31:3 : reserved
+	 * 3 : rx_buf_mirroring
+	 * 31:4 : reserved
 	 */
 	u32 driver_supported_features;
 };
@@ -1129,6 +1130,8 @@ struct ena_admin_ena_mmio_req_read_less_resp {
 #define ENA_ADMIN_HOST_INFO_RX_OFFSET_MASK                  BIT(1)
 #define ENA_ADMIN_HOST_INFO_INTERRUPT_MODERATION_SHIFT      2
 #define ENA_ADMIN_HOST_INFO_INTERRUPT_MODERATION_MASK       BIT(2)
+#define ENA_ADMIN_HOST_INFO_RX_BUF_MIRRORING_SHIFT          3
+#define ENA_ADMIN_HOST_INFO_RX_BUF_MIRRORING_MASK           BIT(3)
 
 /* aenq_common_desc */
 #define ENA_ADMIN_AENQ_COMMON_DESC_PHASE_MASK               BIT(0)
diff --git a/drivers/net/ethernet/amazon/ena/ena_netdev.c b/drivers/net/ethernet/amazon/ena/ena_netdev.c
index 18a30a81a475..fd5f0d87cc59 100644
--- a/drivers/net/ethernet/amazon/ena/ena_netdev.c
+++ b/drivers/net/ethernet/amazon/ena/ena_netdev.c
@@ -959,8 +959,11 @@ static int ena_alloc_rx_page(struct ena_ring *rx_ring,
 		return -ENOMEM;
 	}
 
+	/* To enable NIC-side port-mirroring, AKA SPAN port,
+	 * we make the buffer readable from the nic as well
+	 */
 	dma = dma_map_page(rx_ring->dev, page, 0, ENA_PAGE_SIZE,
-			   DMA_FROM_DEVICE);
+			   DMA_BIDIRECTIONAL);
 	if (unlikely(dma_mapping_error(rx_ring->dev, dma))) {
 		u64_stats_update_begin(&rx_ring->syncp);
 		rx_ring->rx_stats.dma_mapping_err++;
@@ -993,10 +996,9 @@ static void ena_free_rx_page(struct ena_ring *rx_ring,
 		return;
 	}
 
-	dma_unmap_page(rx_ring->dev,
-		       ena_buf->paddr - rx_ring->rx_headroom,
+	dma_unmap_page(rx_ring->dev, ena_buf->paddr - rx_ring->rx_headroom,
 		       ENA_PAGE_SIZE,
-		       DMA_FROM_DEVICE);
+		       DMA_BIDIRECTIONAL);
 
 	__free_page(page);
 	rx_info->page = NULL;
@@ -1431,7 +1433,7 @@ static struct sk_buff *ena_rx_skb(struct ena_ring *rx_ring,
 	do {
 		dma_unmap_page(rx_ring->dev,
 			       dma_unmap_addr(&rx_info->ena_buf, paddr),
-			       ENA_PAGE_SIZE, DMA_FROM_DEVICE);
+			       ENA_PAGE_SIZE, DMA_BIDIRECTIONAL);
 
 		skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, rx_info->page,
 				rx_info->page_offset, len, ENA_PAGE_SIZE);
@@ -3123,7 +3125,8 @@ static void ena_config_host_info(struct ena_com_dev *ena_dev, struct pci_dev *pd
 
 	host_info->driver_supported_features =
 		ENA_ADMIN_HOST_INFO_RX_OFFSET_MASK |
-		ENA_ADMIN_HOST_INFO_INTERRUPT_MODERATION_MASK;
+		ENA_ADMIN_HOST_INFO_INTERRUPT_MODERATION_MASK |
+		ENA_ADMIN_HOST_INFO_RX_BUF_MIRRORING_MASK;
 
 	rc = ena_com_set_host_attributes(ena_dev);
 	if (rc) {
-- 
2.23.1

[PATCH V2 net-next 6/7] net: ena: move llq configuration from ena_probe to ena_device_init()

[akiyano]

From: Arthur Kiyanovski <akiyano@amazon.com>

When the ENA device resets to recover from some error state, all LLQ
configuration values are reset to their defaults, because LLQ is
initialized only once during ena_probe().

Changes in this commit:
1. Move the LLQ configuration process into ena_init_device()
which is called from both ena_probe() and ena_restore_device(). This
way, LLQ setup configurations that are different from the default
values will survive resets.

2. Extract the LLQ bar mapping to ena_map_llq_bar(),
and call once in the lifetime of the driver from ena_probe(),
since there is no need to unmap and map the LLQ bar again every reset.

3. Map the LLQ bar if it exists, regardless if initialization of LLQ
placement policy (ENA_ADMIN_PLACEMENT_POLICY_DEV) succeeded
or not. Initialization might fail the first time, falling back to the
ENA_ADMIN_PLACEMENT_POLICY_HOST placement policy, but later succeed
after device reset, in which case the LLQ bar needs to be mapped
already.

Signed-off-by: Sameeh Jubran <sameehj@amazon.com>
Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
---
 drivers/net/ethernet/amazon/ena/ena_netdev.c | 136 ++++++++++---------
 1 file changed, 73 insertions(+), 63 deletions(-)

diff --git a/drivers/net/ethernet/amazon/ena/ena_netdev.c b/drivers/net/ethernet/amazon/ena/ena_netdev.c
index fd5f0d87cc59..7b8301bac1ce 100644
--- a/drivers/net/ethernet/amazon/ena/ena_netdev.c
+++ b/drivers/net/ethernet/amazon/ena/ena_netdev.c
@@ -3279,10 +3279,71 @@ static int ena_device_validate_params(struct ena_adapter *adapter,
 	return 0;
 }
 
+static void set_default_llq_configurations(struct ena_llq_configurations *llq_config)
+{
+	llq_config->llq_header_location = ENA_ADMIN_INLINE_HEADER;
+	llq_config->llq_stride_ctrl = ENA_ADMIN_MULTIPLE_DESCS_PER_ENTRY;
+	llq_config->llq_num_decs_before_header = ENA_ADMIN_LLQ_NUM_DESCS_BEFORE_HEADER_2;
+	llq_config->llq_ring_entry_size = ENA_ADMIN_LIST_ENTRY_SIZE_128B;
+	llq_config->llq_ring_entry_size_value = 128;
+}
+
+static int ena_set_queues_placement_policy(struct pci_dev *pdev,
+					   struct ena_com_dev *ena_dev,
+					   struct ena_admin_feature_llq_desc *llq,
+					   struct ena_llq_configurations *llq_default_configurations)
+{
+	int rc;
+	u32 llq_feature_mask;
+
+	llq_feature_mask = 1 << ENA_ADMIN_LLQ;
+	if (!(ena_dev->supported_features & llq_feature_mask)) {
+		dev_err(&pdev->dev,
+			"LLQ is not supported Fallback to host mode policy.\n");
+		ena_dev->tx_mem_queue_type = ENA_ADMIN_PLACEMENT_POLICY_HOST;
+		return 0;
+	}
+
+	rc = ena_com_config_dev_mode(ena_dev, llq, llq_default_configurations);
+	if (unlikely(rc)) {
+		dev_err(&pdev->dev,
+			"Failed to configure the device mode.  Fallback to host mode policy.\n");
+		ena_dev->tx_mem_queue_type = ENA_ADMIN_PLACEMENT_POLICY_HOST;
+	}
+
+	return 0;
+}
+
+static int ena_map_llq_mem_bar(struct pci_dev *pdev, struct ena_com_dev *ena_dev,
+			       int bars)
+{
+	bool has_mem_bar = !!(bars & BIT(ENA_MEM_BAR));
+
+	if (!has_mem_bar) {
+		if (ena_dev->tx_mem_queue_type == ENA_ADMIN_PLACEMENT_POLICY_DEV) {
+			dev_err(&pdev->dev,
+				"ENA device does not expose LLQ bar. Fallback to host mode policy.\n");
+			ena_dev->tx_mem_queue_type = ENA_ADMIN_PLACEMENT_POLICY_HOST;
+		}
+
+		return 0;
+	}
+
+	ena_dev->mem_bar = devm_ioremap_wc(&pdev->dev,
+					   pci_resource_start(pdev, ENA_MEM_BAR),
+					   pci_resource_len(pdev, ENA_MEM_BAR));
+
+	if (!ena_dev->mem_bar)
+		return -EFAULT;
+
+	return 0;
+}
+
 static int ena_device_init(struct ena_com_dev *ena_dev, struct pci_dev *pdev,
 			   struct ena_com_dev_get_features_ctx *get_feat_ctx,
 			   bool *wd_state)
 {
+	struct ena_llq_configurations llq_config;
 	struct device *dev = &pdev->dev;
 	bool readless_supported;
 	u32 aenq_groups;
@@ -3373,6 +3434,15 @@ static int ena_device_init(struct ena_com_dev *ena_dev, struct pci_dev *pdev,
 
 	*wd_state = !!(aenq_groups & BIT(ENA_ADMIN_KEEP_ALIVE));
 
+	set_default_llq_configurations(&llq_config);
+
+	rc = ena_set_queues_placement_policy(pdev, ena_dev, &get_feat_ctx->llq,
+					     &llq_config);
+	if (rc) {
+		dev_err(&pdev->dev, "ena device init failed\n");
+		goto err_admin_init;
+	}
+
 	return 0;
 
 err_admin_init:
@@ -3879,54 +3949,6 @@ static u32 ena_calc_max_io_queue_num(struct pci_dev *pdev,
 	return max_num_io_queues;
 }
 
-static int ena_set_queues_placement_policy(struct pci_dev *pdev,
-					   struct ena_com_dev *ena_dev,
-					   struct ena_admin_feature_llq_desc *llq,
-					   struct ena_llq_configurations *llq_default_configurations)
-{
-	bool has_mem_bar;
-	int rc;
-	u32 llq_feature_mask;
-
-	llq_feature_mask = 1 << ENA_ADMIN_LLQ;
-	if (!(ena_dev->supported_features & llq_feature_mask)) {
-		dev_err(&pdev->dev,
-			"LLQ is not supported Fallback to host mode policy.\n");
-		ena_dev->tx_mem_queue_type = ENA_ADMIN_PLACEMENT_POLICY_HOST;
-		return 0;
-	}
-
-	has_mem_bar = pci_select_bars(pdev, IORESOURCE_MEM) & BIT(ENA_MEM_BAR);
-
-	rc = ena_com_config_dev_mode(ena_dev, llq, llq_default_configurations);
-	if (unlikely(rc)) {
-		dev_err(&pdev->dev,
-			"Failed to configure the device mode.  Fallback to host mode policy.\n");
-		ena_dev->tx_mem_queue_type = ENA_ADMIN_PLACEMENT_POLICY_HOST;
-		return 0;
-	}
-
-	/* Nothing to config, exit */
-	if (ena_dev->tx_mem_queue_type == ENA_ADMIN_PLACEMENT_POLICY_HOST)
-		return 0;
-
-	if (!has_mem_bar) {
-		dev_err(&pdev->dev,
-			"ENA device does not expose LLQ bar. Fallback to host mode policy.\n");
-		ena_dev->tx_mem_queue_type = ENA_ADMIN_PLACEMENT_POLICY_HOST;
-		return 0;
-	}
-
-	ena_dev->mem_bar = devm_ioremap_wc(&pdev->dev,
-					   pci_resource_start(pdev, ENA_MEM_BAR),
-					   pci_resource_len(pdev, ENA_MEM_BAR));
-
-	if (!ena_dev->mem_bar)
-		return -EFAULT;
-
-	return 0;
-}
-
 static void ena_set_dev_offloads(struct ena_com_dev_get_features_ctx *feat,
 				 struct net_device *netdev)
 {
@@ -4042,14 +4064,6 @@ static void ena_release_bars(struct ena_com_dev *ena_dev, struct pci_dev *pdev)
 	pci_release_selected_regions(pdev, release_bars);
 }
 
-static void set_default_llq_configurations(struct ena_llq_configurations *llq_config)
-{
-	llq_config->llq_header_location = ENA_ADMIN_INLINE_HEADER;
-	llq_config->llq_ring_entry_size = ENA_ADMIN_LIST_ENTRY_SIZE_128B;
-	llq_config->llq_stride_ctrl = ENA_ADMIN_MULTIPLE_DESCS_PER_ENTRY;
-	llq_config->llq_num_decs_before_header = ENA_ADMIN_LLQ_NUM_DESCS_BEFORE_HEADER_2;
-	llq_config->llq_ring_entry_size_value = 128;
-}
 
 static int ena_calc_io_queue_size(struct ena_calc_queue_size_ctx *ctx)
 {
@@ -4131,7 +4145,6 @@ static int ena_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
 {
 	struct ena_calc_queue_size_ctx calc_queue_ctx = { 0 };
 	struct ena_com_dev_get_features_ctx get_feat_ctx;
-	struct ena_llq_configurations llq_config;
 	struct ena_com_dev *ena_dev = NULL;
 	struct ena_adapter *adapter;
 	struct net_device *netdev;
@@ -4186,13 +4199,10 @@ static int ena_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
 		goto err_free_region;
 	}
 
-	set_default_llq_configurations(&llq_config);
-
-	rc = ena_set_queues_placement_policy(pdev, ena_dev, &get_feat_ctx.llq,
-					     &llq_config);
+	rc = ena_map_llq_mem_bar(pdev, ena_dev, bars);
 	if (rc) {
-		dev_err(&pdev->dev, "ena device init failed\n");
-		goto err_device_destroy;
+		dev_err(&pdev->dev, "ena llq bar mapping failed\n");
+		goto err_free_ena_dev;
 	}
 
 	calc_queue_ctx.ena_dev = ena_dev;
-- 
2.23.1

[PATCH V2 net-next 7/7] net: ena: support new LLQ acceleration mode

[akiyano]

From: Arthur Kiyanovski <akiyano@amazon.com>

New devices add a new hardware acceleration engine, which adds some
restrictions to the driver.
Metadata descriptor must be present for each packet and the maximum
burst size between two doorbells is now limited to a number
advertised by the device.

This patch adds:
1. A handshake protocol between the driver and the device, so the
device will enable the accelerated queues only when both sides
support it.

2. The driver support for the new acceleration engine:
2.1. Send metadata descriptor for each Tx packet.
2.2. Limit the number of packets sent between doorbells.(*)

(*) A previous driver implementation of this feature was comitted in
commit 05d62ca218f8 ("net: ena: add handling of llq max tx burst size")
however the design of the interface between the driver and device
changed since then. This change is reflected in this commit.

Signed-off-by: Netanel Belgazal <netanel@amazon.com>
Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
---
 .../net/ethernet/amazon/ena/ena_admin_defs.h  | 39 ++++++++++++--
 drivers/net/ethernet/amazon/ena/ena_com.c     | 19 ++++++-
 drivers/net/ethernet/amazon/ena/ena_com.h     |  3 ++
 drivers/net/ethernet/amazon/ena/ena_eth_com.c | 51 +++++++++++++------
 drivers/net/ethernet/amazon/ena/ena_eth_com.h |  3 +-
 drivers/net/ethernet/amazon/ena/ena_netdev.c  | 16 ++++--
 drivers/net/ethernet/amazon/ena/ena_netdev.h  |  2 +
 7 files changed, 109 insertions(+), 24 deletions(-)

diff --git a/drivers/net/ethernet/amazon/ena/ena_admin_defs.h b/drivers/net/ethernet/amazon/ena/ena_admin_defs.h
index afe87ff09b20..190dbbde379c 100644
--- a/drivers/net/ethernet/amazon/ena/ena_admin_defs.h
+++ b/drivers/net/ethernet/amazon/ena/ena_admin_defs.h
@@ -491,6 +491,36 @@ enum ena_admin_llq_stride_ctrl {
 	ENA_ADMIN_MULTIPLE_DESCS_PER_ENTRY          = 2,
 };
 
+enum ena_admin_accel_mode_feat {
+	ENA_ADMIN_DISABLE_META_CACHING              = 0,
+	ENA_ADMIN_LIMIT_TX_BURST                    = 1,
+};
+
+struct ena_admin_accel_mode_get {
+	/* bit field of enum ena_admin_accel_mode_feat */
+	u16 supported_flags;
+
+	/* maximum burst size between two doorbells. The size is in bytes */
+	u16 max_tx_burst_size;
+};
+
+struct ena_admin_accel_mode_set {
+	/* bit field of enum ena_admin_accel_mode_feat */
+	u16 enabled_flags;
+
+	u16 reserved;
+};
+
+struct ena_admin_accel_mode_req {
+	union {
+		u32 raw[2];
+
+		struct ena_admin_accel_mode_get get;
+
+		struct ena_admin_accel_mode_set set;
+	} u;
+};
+
 struct ena_admin_feature_llq_desc {
 	u32 max_llq_num;
 
@@ -536,10 +566,13 @@ struct ena_admin_feature_llq_desc {
 	/* the stride control the driver selected to use */
 	u16 descriptors_stride_ctrl_enabled;
 
-	/* Maximum size in bytes taken by llq entries in a single tx burst.
-	 * Set to 0 when there is no such limit.
+	/* reserved */
+	u32 reserved1;
+
+	/* accelerated low latency queues requirement. driver needs to
+	 * support those requirements in order to use accelerated llq
 	 */
-	u32 max_tx_burst_size;
+	struct ena_admin_accel_mode_req accel_mode;
 };
 
 struct ena_admin_queue_ext_feature_fields {
diff --git a/drivers/net/ethernet/amazon/ena/ena_com.c b/drivers/net/ethernet/amazon/ena/ena_com.c
index 432f143559a1..435bf05a853c 100644
--- a/drivers/net/ethernet/amazon/ena/ena_com.c
+++ b/drivers/net/ethernet/amazon/ena/ena_com.c
@@ -403,6 +403,8 @@ static int ena_com_init_io_sq(struct ena_com_dev *ena_dev,
 		       0x0, io_sq->llq_info.desc_list_entry_size);
 		io_sq->llq_buf_ctrl.descs_left_in_line =
 			io_sq->llq_info.descs_num_before_header;
+		io_sq->disable_meta_caching =
+			io_sq->llq_info.disable_meta_caching;
 
 		if (io_sq->llq_info.max_entries_in_tx_burst > 0)
 			io_sq->entries_in_tx_burst_left =
@@ -626,6 +628,10 @@ static int ena_com_set_llq(struct ena_com_dev *ena_dev)
 	cmd.u.llq.desc_num_before_header_enabled = llq_info->descs_num_before_header;
 	cmd.u.llq.descriptors_stride_ctrl_enabled = llq_info->desc_stride_ctrl;
 
+	cmd.u.llq.accel_mode.u.set.enabled_flags =
+		BIT(ENA_ADMIN_DISABLE_META_CACHING) |
+		BIT(ENA_ADMIN_LIMIT_TX_BURST);
+
 	ret = ena_com_execute_admin_command(admin_queue,
 					    (struct ena_admin_aq_entry *)&cmd,
 					    sizeof(cmd),
@@ -643,6 +649,7 @@ static int ena_com_config_llq_info(struct ena_com_dev *ena_dev,
 				   struct ena_llq_configurations *llq_default_cfg)
 {
 	struct ena_com_llq_info *llq_info = &ena_dev->llq_info;
+	struct ena_admin_accel_mode_get llq_accel_mode_get;
 	u16 supported_feat;
 	int rc;
 
@@ -742,9 +749,17 @@ static int ena_com_config_llq_info(struct ena_com_dev *ena_dev,
 		       llq_default_cfg->llq_num_decs_before_header,
 		       supported_feat, llq_info->descs_num_before_header);
 	}
+	/* Check for accelerated queue supported */
+	llq_accel_mode_get = llq_features->accel_mode.u.get;
+
+	llq_info->disable_meta_caching =
+		!!(llq_accel_mode_get.supported_flags &
+		   BIT(ENA_ADMIN_DISABLE_META_CACHING));
 
-	llq_info->max_entries_in_tx_burst =
-		(u16)(llq_features->max_tx_burst_size /	llq_default_cfg->llq_ring_entry_size_value);
+	if (llq_accel_mode_get.supported_flags & BIT(ENA_ADMIN_LIMIT_TX_BURST))
+		llq_info->max_entries_in_tx_burst =
+			llq_accel_mode_get.max_tx_burst_size /
+			llq_default_cfg->llq_ring_entry_size_value;
 
 	rc = ena_com_set_llq(ena_dev);
 	if (rc)
diff --git a/drivers/net/ethernet/amazon/ena/ena_com.h b/drivers/net/ethernet/amazon/ena/ena_com.h
index 4c98f6f07882..4287d47b2b0b 100644
--- a/drivers/net/ethernet/amazon/ena/ena_com.h
+++ b/drivers/net/ethernet/amazon/ena/ena_com.h
@@ -127,6 +127,7 @@ struct ena_com_llq_info {
 	u16 descs_num_before_header;
 	u16 descs_per_entry;
 	u16 max_entries_in_tx_burst;
+	bool disable_meta_caching;
 };
 
 struct ena_com_io_cq {
@@ -189,6 +190,8 @@ struct ena_com_io_sq {
 	enum queue_direction direction;
 	enum ena_admin_placement_policy_type mem_queue_type;
 
+	bool disable_meta_caching;
+
 	u32 msix_vector;
 	struct ena_com_tx_meta cached_tx_meta;
 	struct ena_com_llq_info llq_info;
diff --git a/drivers/net/ethernet/amazon/ena/ena_eth_com.c b/drivers/net/ethernet/amazon/ena/ena_eth_com.c
index ec8ea25e988d..ccd440589565 100644
--- a/drivers/net/ethernet/amazon/ena/ena_eth_com.c
+++ b/drivers/net/ethernet/amazon/ena/ena_eth_com.c
@@ -285,11 +285,10 @@ static u16 ena_com_cdesc_rx_pkt_get(struct ena_com_io_cq *io_cq,
 	return count;
 }
 
-static int ena_com_create_and_store_tx_meta_desc(struct ena_com_io_sq *io_sq,
-							struct ena_com_tx_ctx *ena_tx_ctx)
+static int ena_com_create_meta(struct ena_com_io_sq *io_sq,
+			       struct ena_com_tx_meta *ena_meta)
 {
 	struct ena_eth_io_tx_meta_desc *meta_desc = NULL;
-	struct ena_com_tx_meta *ena_meta = &ena_tx_ctx->ena_meta;
 
 	meta_desc = get_sq_desc(io_sq);
 	memset(meta_desc, 0x0, sizeof(struct ena_eth_io_tx_meta_desc));
@@ -309,12 +308,13 @@ static int ena_com_create_and_store_tx_meta_desc(struct ena_com_io_sq *io_sq,
 
 	/* Extended meta desc */
 	meta_desc->len_ctrl |= ENA_ETH_IO_TX_META_DESC_ETH_META_TYPE_MASK;
-	meta_desc->len_ctrl |= ENA_ETH_IO_TX_META_DESC_META_STORE_MASK;
 	meta_desc->len_ctrl |= (io_sq->phase <<
 		ENA_ETH_IO_TX_META_DESC_PHASE_SHIFT) &
 		ENA_ETH_IO_TX_META_DESC_PHASE_MASK;
 
 	meta_desc->len_ctrl |= ENA_ETH_IO_TX_META_DESC_FIRST_MASK;
+	meta_desc->len_ctrl |= ENA_ETH_IO_TX_META_DESC_META_STORE_MASK;
+
 	meta_desc->word2 |= ena_meta->l3_hdr_len &
 		ENA_ETH_IO_TX_META_DESC_L3_HDR_LEN_MASK;
 	meta_desc->word2 |= (ena_meta->l3_hdr_offset <<
@@ -325,13 +325,36 @@ static int ena_com_create_and_store_tx_meta_desc(struct ena_com_io_sq *io_sq,
 		ENA_ETH_IO_TX_META_DESC_L4_HDR_LEN_IN_WORDS_SHIFT) &
 		ENA_ETH_IO_TX_META_DESC_L4_HDR_LEN_IN_WORDS_MASK;
 
-	meta_desc->len_ctrl |= ENA_ETH_IO_TX_META_DESC_META_STORE_MASK;
+	return ena_com_sq_update_tail(io_sq);
+}
+
+static int ena_com_create_and_store_tx_meta_desc(struct ena_com_io_sq *io_sq,
+						 struct ena_com_tx_ctx *ena_tx_ctx,
+						 bool *have_meta)
+{
+	struct ena_com_tx_meta *ena_meta = &ena_tx_ctx->ena_meta;
 
-	/* Cached the meta desc */
-	memcpy(&io_sq->cached_tx_meta, ena_meta,
-	       sizeof(struct ena_com_tx_meta));
+	/* When disable meta caching is set, don't bother to save the meta and
+	 * compare it to the stored version, just create the meta
+	 */
+	if (io_sq->disable_meta_caching) {
+		if (unlikely(!ena_tx_ctx->meta_valid))
+			return -EINVAL;
 
-	return ena_com_sq_update_tail(io_sq);
+		*have_meta = true;
+		return ena_com_create_meta(io_sq, ena_meta);
+	}
+
+	if (ena_com_meta_desc_changed(io_sq, ena_tx_ctx)) {
+		*have_meta = true;
+		/* Cache the meta desc */
+		memcpy(&io_sq->cached_tx_meta, ena_meta,
+		       sizeof(struct ena_com_tx_meta));
+		return ena_com_create_meta(io_sq, ena_meta);
+	}
+
+	*have_meta = false;
+	return 0;
 }
 
 static void ena_com_rx_set_flags(struct ena_com_rx_ctx *ena_rx_ctx,
@@ -402,12 +425,10 @@ int ena_com_prepare_tx(struct ena_com_io_sq *io_sq,
 	if (unlikely(rc))
 		return rc;
 
-	have_meta = ena_tx_ctx->meta_valid && ena_com_meta_desc_changed(io_sq,
-			ena_tx_ctx);
-	if (have_meta) {
-		rc = ena_com_create_and_store_tx_meta_desc(io_sq, ena_tx_ctx);
-		if (unlikely(rc))
-			return rc;
+	rc = ena_com_create_and_store_tx_meta_desc(io_sq, ena_tx_ctx, &have_meta);
+	if (unlikely(rc)) {
+		pr_err("failed to create and store tx meta desc\n");
+		return rc;
 	}
 
 	/* If the caller doesn't want to send packets */
diff --git a/drivers/net/ethernet/amazon/ena/ena_eth_com.h b/drivers/net/ethernet/amazon/ena/ena_eth_com.h
index 8b1afd3b32f2..b6592cb93b04 100644
--- a/drivers/net/ethernet/amazon/ena/ena_eth_com.h
+++ b/drivers/net/ethernet/amazon/ena/ena_eth_com.h
@@ -157,7 +157,8 @@ static inline bool ena_com_is_doorbell_needed(struct ena_com_io_sq *io_sq,
 	llq_info = &io_sq->llq_info;
 	num_descs = ena_tx_ctx->num_bufs;
 
-	if (unlikely(ena_com_meta_desc_changed(io_sq, ena_tx_ctx)))
+	if (llq_info->disable_meta_caching ||
+	    unlikely(ena_com_meta_desc_changed(io_sq, ena_tx_ctx)))
 		++num_descs;
 
 	if (num_descs > llq_info->descs_num_before_header) {
diff --git a/drivers/net/ethernet/amazon/ena/ena_netdev.c b/drivers/net/ethernet/amazon/ena/ena_netdev.c
index 7b8301bac1ce..455fb4c6b4cf 100644
--- a/drivers/net/ethernet/amazon/ena/ena_netdev.c
+++ b/drivers/net/ethernet/amazon/ena/ena_netdev.c
@@ -655,6 +655,7 @@ static void ena_init_io_rings(struct ena_adapter *adapter,
 		txr->sgl_size = adapter->max_tx_sgl_size;
 		txr->smoothed_interval =
 			ena_com_get_nonadaptive_moderation_interval_tx(ena_dev);
+		txr->disable_meta_caching = adapter->disable_meta_caching;
 
 		/* Don't init RX queues for xdp queues */
 		if (!ENA_IS_XDP_INDEX(adapter, i)) {
@@ -2783,7 +2784,9 @@ int ena_update_queue_count(struct ena_adapter *adapter, u32 new_channel_count)
 	return dev_was_up ? ena_open(adapter->netdev) : 0;
 }
 
-static void ena_tx_csum(struct ena_com_tx_ctx *ena_tx_ctx, struct sk_buff *skb)
+static void ena_tx_csum(struct ena_com_tx_ctx *ena_tx_ctx,
+			struct sk_buff *skb,
+			bool disable_meta_caching)
 {
 	u32 mss = skb_shinfo(skb)->gso_size;
 	struct ena_com_tx_meta *ena_meta = &ena_tx_ctx->ena_meta;
@@ -2827,7 +2830,9 @@ static void ena_tx_csum(struct ena_com_tx_ctx *ena_tx_ctx, struct sk_buff *skb)
 		ena_meta->l3_hdr_len = skb_network_header_len(skb);
 		ena_meta->l3_hdr_offset = skb_network_offset(skb);
 		ena_tx_ctx->meta_valid = 1;
-
+	} else if (disable_meta_caching) {
+		memset(ena_meta, 0, sizeof(*ena_meta));
+		ena_tx_ctx->meta_valid = 1;
 	} else {
 		ena_tx_ctx->meta_valid = 0;
 	}
@@ -3011,7 +3016,7 @@ static netdev_tx_t ena_start_xmit(struct sk_buff *skb, struct net_device *dev)
 	ena_tx_ctx.header_len = header_len;
 
 	/* set flags and meta data */
-	ena_tx_csum(&ena_tx_ctx, skb);
+	ena_tx_csum(&ena_tx_ctx, skb, tx_ring->disable_meta_caching);
 
 	rc = ena_xmit_common(dev,
 			     tx_ring,
@@ -4259,6 +4264,11 @@ static int ena_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
 	adapter->xdp_num_queues = 0;
 
 	adapter->rx_copybreak = ENA_DEFAULT_RX_COPYBREAK;
+	if (ena_dev->tx_mem_queue_type == ENA_ADMIN_PLACEMENT_POLICY_DEV)
+		adapter->disable_meta_caching =
+			!!(get_feat_ctx.llq.accel_mode.u.get.supported_flags &
+			   BIT(ENA_ADMIN_DISABLE_META_CACHING));
+
 	adapter->wd_state = wd_state;
 
 	snprintf(adapter->name, ENA_NAME_MAX_LEN, "ena_%d", adapters_found);
diff --git a/drivers/net/ethernet/amazon/ena/ena_netdev.h b/drivers/net/ethernet/amazon/ena/ena_netdev.h
index 89304b403995..0c8504006247 100644
--- a/drivers/net/ethernet/amazon/ena/ena_netdev.h
+++ b/drivers/net/ethernet/amazon/ena/ena_netdev.h
@@ -298,6 +298,7 @@ struct ena_ring {
 	u8 tx_max_header_size;
 
 	bool first_interrupt;
+	bool disable_meta_caching;
 	u16 no_interrupt_event_cnt;
 
 	/* cpu for TPH */
@@ -399,6 +400,7 @@ struct ena_adapter {
 
 	bool wd_state;
 	bool dev_up_before_reset;
+	bool disable_meta_caching;
 	unsigned long last_keep_alive_jiffies;
 
 	struct u64_stats_sync syncp;
-- 
2.23.1

Re: [PATCH V2 net-next 1/7] net: ena: avoid unnecessary rearming of interrupt vector when busy-polling

[Eric Dumazet]

On 7/12/20 3:36 PM, akiyano@amazon.com wrote:
> From: Arthur Kiyanovski <akiyano@amazon.com>
> 
> In napi busy-poll mode, the kernel invokes the napi handler of the
> device repeatedly to poll the NIC's receive queues. This process
> repeats until a timeout, specific for each connection, is up.
> By polling packets in busy-poll mode the user may gain lower latency
> and higher throughput (since the kernel no longer waits for interrupts
> to poll the queues) in expense of CPU usage.
> 
> Upon completing a napi routine, the driver checks whether
> the routine was called by an interrupt handler. If so, the driver
> re-enables interrupts for the device. This is needed since an
> interrupt routine invocation disables future invocations until
> explicitly re-enabled.
> 
> The driver avoids re-enabling the interrupts if they were not disabled
> in the first place (e.g. if driver in busy mode).
> Originally, the driver checked whether interrupt re-enabling is needed
> by reading the 'ena_napi->unmask_interrupt' variable. This atomic
> variable was set upon interrupt and cleared after re-enabling it.
> 
> In the 4.10 Linux version, the 'napi_complete_done' call was changed
> so that it returns 'false' when device should not re-enable
> interrupts, and 'true' otherwise. The change includes reading the
> "NAPIF_STATE_IN_BUSY_POLL" flag to check if the napi call is in
> busy-poll mode, and if so, return 'false'.
> The driver was changed to re-enable interrupts according to this
> routine's return value.
> The Linux community rejected the use of the
> 'ena_napi->unmaunmask_interrupt' variable to determine whether
> unmasking is needed, and urged to use napi_napi_complete_done()
> return value solely.
> See https://lore.kernel.org/patchwork/patch/741149/ for more details

Yeah, and I see you did not bother to CC me on this new submission.

> 
> As explained, a busy-poll session exists for a specified timeout
> value, after which it exits the busy-poll mode and re-enters it later.
> This leads to many invocations of the napi handler where
> napi_complete_done() false indicates that interrupts should be
> re-enabled.
> This creates a bug in which the interrupts are re-enabled
> unnecessarily.
> To reproduce this bug:
>     1) echo 50 | sudo tee /proc/sys/net/core/busy_poll
>     2) echo 50 | sudo tee /proc/sys/net/core/busy_read
>     3) Add counters that check whether
>     'ena_unmask_interrupt(tx_ring, rx_ring);'
>     is called without disabling the interrupts in the first
>     place (i.e. with calling the interrupt routine
>     ena_intr_msix_io())
> 
> Steps 1+2 enable busy-poll as the default mode for new connections.
> 
> The busy poll routine rearms the interrupts after every session by
> design, and so we need to add an extra check that the interrupts were
> masked in the first place.
> 
> Signed-off-by: Shay Agroskin <shayagr@amazon.com>
> Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
> ---
>  drivers/net/ethernet/amazon/ena/ena_netdev.c | 7 ++++++-
>  drivers/net/ethernet/amazon/ena/ena_netdev.h | 1 +
>  2 files changed, 7 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/net/ethernet/amazon/ena/ena_netdev.c b/drivers/net/ethernet/amazon/ena/ena_netdev.c
> index 91be3ffa1c5c..90c0fe15cd23 100644
> --- a/drivers/net/ethernet/amazon/ena/ena_netdev.c
> +++ b/drivers/net/ethernet/amazon/ena/ena_netdev.c
> @@ -1913,7 +1913,9 @@ static int ena_io_poll(struct napi_struct *napi, int budget)
>  		/* Update numa and unmask the interrupt only when schedule
>  		 * from the interrupt context (vs from sk_busy_loop)
>  		 */
> -		if (napi_complete_done(napi, rx_work_done)) {
> +		if (napi_complete_done(napi, rx_work_done) &&
> +		    READ_ONCE(ena_napi->interrupts_masked)) {
> +			WRITE_ONCE(ena_napi->interrupts_masked, false);
>  			/* We apply adaptive moderation on Rx path only.
>  			 * Tx uses static interrupt moderation.
>  			 */
> @@ -1961,6 +1963,9 @@ static irqreturn_t ena_intr_msix_io(int irq, void *data)
>  
>  	ena_napi->first_interrupt = true;
>  
> +	WRITE_ONCE(ena_napi->interrupts_masked, true);
> +	smp_wmb(); /* write interrupts_masked before calling napi */

It is not clear where is the paired smp_wmb() 

> +
>  	napi_schedule_irqoff(&ena_napi->napi);
>  
>  	return IRQ_HANDLED;
> diff --git a/drivers/net/ethernet/amazon/ena/ena_netdev.h b/drivers/net/ethernet/amazon/ena/ena_netdev.h
> index ba030d260940..89304b403995 100644
> --- a/drivers/net/ethernet/amazon/ena/ena_netdev.h
> +++ b/drivers/net/ethernet/amazon/ena/ena_netdev.h
> @@ -167,6 +167,7 @@ struct ena_napi {
>  	struct ena_ring *rx_ring;
>  	struct ena_ring *xdp_ring;
>  	bool first_interrupt;
> +	bool interrupts_masked;
>  	u32 qid;
>  	struct dim dim;
>  };
> 

Note that writing/reading over bool fields from hard irq context without proper sync is not generally allowed.

Compiler could perform RMW over plain 32bit words.

Sometimes we accept the races, but in this context this might be bad.

Re: [PATCH V2 net-next 1/7] net: ena: avoid unnecessary rearming of interrupt vector when busy-polling

[Shay Agroskin]

Eric Dumazet <eric.dumazet@gmail.com> writes:

> CAUTION: This email originated from outside of the organization. Do not click links or open attachments unless you can confirm the sender and know the content is safe.
>
>
>
> On 7/12/20 3:36 PM, akiyano@amazon.com wrote:
>> From: Arthur Kiyanovski <akiyano@amazon.com>
>>
>> In napi busy-poll mode, the kernel invokes the napi handler of the
>> device repeatedly to poll the NIC's receive queues. This process
>> repeats until a timeout, specific for each connection, is up.
>> By polling packets in busy-poll mode the user may gain lower latency
>> and higher throughput (since the kernel no longer waits for interrupts
>> to poll the queues) in expense of CPU usage.
>>
>> Upon completing a napi routine, the driver checks whether
>> the routine was called by an interrupt handler. If so, the driver
>> re-enables interrupts for the device. This is needed since an
>> interrupt routine invocation disables future invocations until
>> explicitly re-enabled.
>>
>> The driver avoids re-enabling the interrupts if they were not disabled
>> in the first place (e.g. if driver in busy mode).
>> Originally, the driver checked whether interrupt re-enabling is needed
>> by reading the 'ena_napi->unmask_interrupt' variable. This atomic
>> variable was set upon interrupt and cleared after re-enabling it.
>>
>> In the 4.10 Linux version, the 'napi_complete_done' call was changed
>> so that it returns 'false' when device should not re-enable
>> interrupts, and 'true' otherwise. The change includes reading the
>> "NAPIF_STATE_IN_BUSY_POLL" flag to check if the napi call is in
>> busy-poll mode, and if so, return 'false'.
>> The driver was changed to re-enable interrupts according to this
>> routine's return value.
>> The Linux community rejected the use of the
>> 'ena_napi->unmaunmask_interrupt' variable to determine whether
>> unmasking is needed, and urged to use napi_napi_complete_done()
>> return value solely.
>> See https://lore.kernel.org/patchwork/patch/741149/ for more details
>
> Yeah, and I see you did not bother to CC me on this new submission.
>

I apologize for this. We should have been more careful. We'll take
better notice next time

>>
>> As explained, a busy-poll session exists for a specified timeout
>> value, after which it exits the busy-poll mode and re-enters it later.
>> This leads to many invocations of the napi handler where
>> napi_complete_done() false indicates that interrupts should be
>> re-enabled.
>> This creates a bug in which the interrupts are re-enabled
>> unnecessarily.
>> To reproduce this bug:
>>     1) echo 50 | sudo tee /proc/sys/net/core/busy_poll
>>     2) echo 50 | sudo tee /proc/sys/net/core/busy_read
>>     3) Add counters that check whether
>>     'ena_unmask_interrupt(tx_ring, rx_ring);'
>>     is called without disabling the interrupts in the first
>>     place (i.e. with calling the interrupt routine
>>     ena_intr_msix_io())
>>
>> Steps 1+2 enable busy-poll as the default mode for new connections.
>>
>> The busy poll routine rearms the interrupts after every session by
>> design, and so we need to add an extra check that the interrupts were
>> masked in the first place.
>>
>> Signed-off-by: Shay Agroskin <shayagr@amazon.com>
>> Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
>> ---
>>  drivers/net/ethernet/amazon/ena/ena_netdev.c | 7 ++++++-
>>  drivers/net/ethernet/amazon/ena/ena_netdev.h | 1 +
>>  2 files changed, 7 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/net/ethernet/amazon/ena/ena_netdev.c b/drivers/net/ethernet/amazon/ena/ena_netdev.c
>> index 91be3ffa1c5c..90c0fe15cd23 100644
>> --- a/drivers/net/ethernet/amazon/ena/ena_netdev.c
>> +++ b/drivers/net/ethernet/amazon/ena/ena_netdev.c
>> @@ -1913,7 +1913,9 @@ static int ena_io_poll(struct napi_struct *napi, int budget)
>>               /* Update numa and unmask the interrupt only when schedule
>>                * from the interrupt context (vs from sk_busy_loop)
>>                */
>> -             if (napi_complete_done(napi, rx_work_done)) {
>> +             if (napi_complete_done(napi, rx_work_done) &&
>> +                 READ_ONCE(ena_napi->interrupts_masked)) {
>> +                     WRITE_ONCE(ena_napi->interrupts_masked, false);
>>                       /* We apply adaptive moderation on Rx path only.
>>                        * Tx uses static interrupt moderation.
>>                        */
>> @@ -1961,6 +1963,9 @@ static irqreturn_t ena_intr_msix_io(int irq, void *data)
>>
>>       ena_napi->first_interrupt = true;
>>
>> +     WRITE_ONCE(ena_napi->interrupts_masked, true);
>> +     smp_wmb(); /* write interrupts_masked before calling napi */
>
> It is not clear where is the paired smp_wmb()
>
Can you please explain what you mean ? The idea of adding the store barrier here is to ensure that the WRITE_ONCE(…) invocation is executed before
invoking the napi soft irq. From what I gathered using this command would result in compiler barrier (which would prevent it from executing the bool store after napi scheduling) on x86
and a memory barrier on ARM64 machines which have a weaker consistency model.
>> +
>>       napi_schedule_irqoff(&ena_napi->napi);
>>
>>       return IRQ_HANDLED;
>> diff --git a/drivers/net/ethernet/amazon/ena/ena_netdev.h b/drivers/net/ethernet/amazon/ena/ena_netdev.h
>> index ba030d260940..89304b403995 100644
>> --- a/drivers/net/ethernet/amazon/ena/ena_netdev.h
>> +++ b/drivers/net/ethernet/amazon/ena/ena_netdev.h
>> @@ -167,6 +167,7 @@ struct ena_napi {
>>       struct ena_ring *rx_ring;
>>       struct ena_ring *xdp_ring;
>>       bool first_interrupt;
>> +     bool interrupts_masked;
>>       u32 qid;
>>       struct dim dim;
>>  };
>>
>
> Note that writing/reading over bool fields from hard irq context without proper sync is not generally allowed.
>
> Compiler could perform RMW over plain 32bit words.

Doesn't the READ/WRITE_ONCE macros prevent the compiler from reordering these instructions ? Also from what I researched (please correct me if I'm wrong here)
both x86 and ARM don't allow reordering LOAD and STORE when they access
the same variable (register or memory address).
>
> Sometimes we accept the races, but in this context this might be bad.

As long a the writing of the flag is atomic in the sense that the value in memory isn't some hybrid value of two parallel stores, the existing race cannot result in a dead lock or
leaving the interrupt vector masked. Am I missing something ?

Thank you for taking the time to look at this patch.

Shay

Re: [PATCH V2 net-next 1/7] net: ena: avoid unnecessary rearming of interrupt vector when busy-polling

[Eric Dumazet]

On 7/13/20 12:39 PM, Shay Agroskin wrote:
> 
> Eric Dumazet <eric.dumazet@gmail.com> writes:
> 

>>> +     WRITE_ONCE(ena_napi->interrupts_masked, true);
>>> +     smp_wmb(); /* write interrupts_masked before calling napi */
>>
>> It is not clear where is the paired smp_wmb()
>>
> Can you please explain what you mean ? The idea of adding the store barrier here is to ensure that the WRITE_ONCE(…) invocation is executed before
> invoking the napi soft irq. From what I gathered using this command would result in compiler barrier (which would prevent it from executing the bool store after napi scheduling) on x86
> and a memory barrier on ARM64 machines which have a weaker consistency model.

Every time you add a smp_wmb() somewhere, the question is raised where the opposite barrier (usually smp_rmb())
is used.

You should document this, pointing where is the opposite smp_rmb()

If you can not find it (READ_ONCE() has no implied smp_rmb()), then
something might be wrong in your patch.

Re: [PATCH V2 net-next 1/7] net: ena: avoid unnecessary rearming of interrupt vector when busy-polling

[Bshara, Nafea]

    >>
    >> As explained, a busy-poll session exists for a specified timeout
    >> value, after which it exits the busy-poll mode and re-enters it later.
    >> This leads to many invocations of the napi handler where
    >> napi_complete_done() false indicates that interrupts should be
    >> re-enabled.
    >> This creates a bug in which the interrupts are re-enabled
    >> unnecessarily.
    >> To reproduce this bug:
    >>     1) echo 50 | sudo tee /proc/sys/net/core/busy_poll
    >>     2) echo 50 | sudo tee /proc/sys/net/core/busy_read
    >>     3) Add counters that check whether
    >>     'ena_unmask_interrupt(tx_ring, rx_ring);'
    >>     is called without disabling the interrupts in the first
    >>     place (i.e. with calling the interrupt routine
    >>     ena_intr_msix_io())
    >>
    >> Steps 1+2 enable busy-poll as the default mode for new connections.
    >>
    >> The busy poll routine rearms the interrupts after every session by
    >> design, and so we need to add an extra check that the interrupts were
    >> masked in the first place.
    >>
    >> Signed-off-by: Shay Agroskin <shayagr@amazon.com>
    >> Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
    >> ---
    >>  drivers/net/ethernet/amazon/ena/ena_netdev.c | 7 ++++++-
    >>  drivers/net/ethernet/amazon/ena/ena_netdev.h | 1 +
    >>  2 files changed, 7 insertions(+), 1 deletion(-)
    >>
    >> diff --git a/drivers/net/ethernet/amazon/ena/ena_netdev.c b/drivers/net/ethernet/amazon/ena/ena_netdev.c
    >> index 91be3ffa1c5c..90c0fe15cd23 100644
    >> --- a/drivers/net/ethernet/amazon/ena/ena_netdev.c
    >> +++ b/drivers/net/ethernet/amazon/ena/ena_netdev.c
    >> @@ -1913,7 +1913,9 @@ static int ena_io_poll(struct napi_struct *napi, int budget)
    >>               /* Update numa and unmask the interrupt only when schedule
    >>                * from the interrupt context (vs from sk_busy_loop)
    >>                */
    >> -             if (napi_complete_done(napi, rx_work_done)) {
    >> +             if (napi_complete_done(napi, rx_work_done) &&
    >> +                 READ_ONCE(ena_napi->interrupts_masked)) {
    >> +                     WRITE_ONCE(ena_napi->interrupts_masked, false);
    >>                       /* We apply adaptive moderation on Rx path only.
    >>                        * Tx uses static interrupt moderation.
    >>                        */
    >> @@ -1961,6 +1963,9 @@ static irqreturn_t ena_intr_msix_io(int irq, void *data)
    >>
    >>       ena_napi->first_interrupt = true;
    >>
    >> +     WRITE_ONCE(ena_napi->interrupts_masked, true);
    >> +     smp_wmb(); /* write interrupts_masked before calling napi */
    >
    > It is not clear where is the paired smp_wmb()
    >
    Can you please explain what you mean ? The idea of adding the store barrier here is to ensure that the WRITE_ONCE(…) invocation is executed before

[NB] There are two aspects .  if we doing smp_wmb() when WRITE_ONCE(...true),  then u need to so smp_wmb() in the place u do WRITE_ONCE(...false)

[NB] Eric also highlighted need for smp_rmb().  That's not needed here in my opinion
[NB] as the main objective is to make the write observable across all the cores in CPUs that have weaker consistency model and don’t guarantee write observability across all cores (like arm and ppc)

    invoking the napi soft irq. From what I gathered using this command would result in compiler barrier (which would prevent it from executing the bool store after napi scheduling) on x86
    and a memory barrier on ARM64 machines which have a weaker consistency model.
    >> +
    >>       napi_schedule_irqoff(&ena_napi->napi);
    >>
    >>       return IRQ_HANDLED;
    >> diff --git a/drivers/net/ethernet/amazon/ena/ena_netdev.h b/drivers/net/ethernet/amazon/ena/ena_netdev.h
    >> index ba030d260940..89304b403995 100644
    >> --- a/drivers/net/ethernet/amazon/ena/ena_netdev.h
    >> +++ b/drivers/net/ethernet/amazon/ena/ena_netdev.h
    >> @@ -167,6 +167,7 @@ struct ena_napi {
    >>       struct ena_ring *rx_ring;
    >>       struct ena_ring *xdp_ring;
    >>       bool first_interrupt;
    >> +     bool interrupts_masked;
    >>       u32 qid;
    >>       struct dim dim;
    >>  };
    >>
    >
    > Note that writing/reading over bool fields from hard irq context without proper sync is not generally allowed.
    >
    > Compiler could perform RMW over plain 32bit words.

    Doesn't the READ/WRITE_ONCE macros prevent the compiler from reordering these instructions ? Also from what I researched (please correct me if I'm wrong here)
    both x86 and ARM don't allow reordering LOAD and STORE when they access
    the same variable (register or memory address).

[NB] that is true within the same core.  But if store is in interrupt routine, and load is in napi, they could be running on different cores hence you use smp_wmb to make it observable
[NB] the key in this design that u set the bit, send smp_wmb(), before waking up napi, or ordering and observability is guaranteed

    >
    > Sometimes we accept the races, but in this context this might be bad.

    As long a the writing of the flag is atomic in the sense that the value in memory isn't some hybrid value of two parallel stores, the existing race cannot result in a dead lock or
    leaving the interrupt vector masked. Am I missing something ?

[NB] the race would exist if napi was running in same time interrupt routine is running
[NB] but in ENA design, that wont happen, and it is guarantee that only one of them is running at same time, as the interrupt is unmasked only at the end of napi() job
[NB] as Eric mention, this should be documented

    Thank you for taking the time to look at this patch.

    Shay