[PATCH net-next 0/2] net: macb: implement ethtool set channels count operation

[PATCH net-next 0/2] net: macb: implement ethtool set channels count operation

[Théo Lebrun]

Add support for changing the active number of queues. Tested on Mobileye
EyeQ5. The first patch is as expected. However the second one might be
more surprising:

GEM has per-queue Tx SRAM segmentation. If we do not touch SRAM
distribution then we'll only be able to exploit a portion of it when a
smaller queue count is configured. It also is beneficial if bootloader
stages write to the register and we don't reset it but attempt to use
all queues (the default).

The operation is only hidden behind MACB_CAPS_QUEUE_DISABLE, we do not
introduce yet another feature flag.

As this series must s/num_queues/max_num_queues/ quite a lot, it
conflicts with other in-flights series touching MACB in the netdev ML.
I'll make sure to resend once net-next gets updated.

Have a nice day,
Thanks,
Théo

Signed-off-by: Théo Lebrun <theo.lebrun@bootlin.com>
---
Théo Lebrun (2):
      net: macb: implement ethtool_ops.get|set_channels()
      net: macb: distribute evenly Tx SRAM segments

 drivers/net/ethernet/cadence/Kconfig     |   6 +
 drivers/net/ethernet/cadence/macb.h      |   6 +
 drivers/net/ethernet/cadence/macb_main.c | 235 ++++++++++++++++++++++++++++---
 3 files changed, 231 insertions(+), 16 deletions(-)
---
base-commit: 3d0508701c2bff577d2ce4fe73bd8954fd969f43
change-id: 20260305-macb-set-channels-5bf6e07f3270

Best regards,
-- 
Théo Lebrun <theo.lebrun@bootlin.com>

[PATCH net-next 1/2] net: macb: implement ethtool_ops.get|set_channels()

[Théo Lebrun]

bp->num_queues is the total number of queues and is constant from probe.
Introduce bp->max_num_queues which takes the current role of
bp->num_queues and allow `0 < bp->num_queues <= bp->max_num_queues`.
MACB/GEM does not know about rx/tx specific queues; it only has
combined queues.

Implement .set_channels() operation by wrapping ourselves in
macb_close() and macb_open() calls if interface is running. This
triggers reinit of buffers, which also includes the code that
enables/disables only the queues below bp->num_queues, in
macb_init_buffers().

If reopen fails, then we do one last attempt and reset num_queues to
the previous value and try reopen. We still error out because the
.set_channels() operation failed.

.set_channels() is reserved to devices with MACB_CAPS_QUEUE_DISABLE.
The tieoff workaround would not work as packets would still be routed
into queues with a tieoff descriptor.

Nit: fix an alignment issue inside gem_ethtool_ops which does not
deserve its own patch.

Signed-off-by: Théo Lebrun <theo.lebrun@bootlin.com>
---
 drivers/net/ethernet/cadence/macb.h      |   1 +
 drivers/net/ethernet/cadence/macb_main.c | 100 ++++++++++++++++++++++++++-----
 2 files changed, 85 insertions(+), 16 deletions(-)

diff --git a/drivers/net/ethernet/cadence/macb.h b/drivers/net/ethernet/cadence/macb.h
index 87414a2ddf6e..30fa65e2bdf2 100644
--- a/drivers/net/ethernet/cadence/macb.h
+++ b/drivers/net/ethernet/cadence/macb.h
@@ -1294,6 +1294,7 @@ struct macb {
 	unsigned int		tx_ring_size;
 
 	unsigned int		num_queues;
+	unsigned int		max_num_queues;
 	struct macb_queue	queues[MACB_MAX_QUEUES];
 
 	spinlock_t		lock;
diff --git a/drivers/net/ethernet/cadence/macb_main.c b/drivers/net/ethernet/cadence/macb_main.c
index 17f0ad3d7a09..bac83a2b4c4d 100644
--- a/drivers/net/ethernet/cadence/macb_main.c
+++ b/drivers/net/ethernet/cadence/macb_main.c
@@ -466,9 +466,24 @@ static void macb_init_buffers(struct macb *bp)
 			    upper_32_bits(bp->queues[0].tx_ring_dma));
 	}
 
-	for (q = 0, queue = bp->queues; q < bp->num_queues; ++q, ++queue) {
-		queue_writel(queue, RBQP, lower_32_bits(queue->rx_ring_dma));
-		queue_writel(queue, TBQP, lower_32_bits(queue->tx_ring_dma));
+	for (q = 0, queue = bp->queues; q < bp->max_num_queues; ++q, ++queue) {
+		if (q < bp->num_queues) {
+			queue_writel(queue, RBQP, lower_32_bits(queue->rx_ring_dma));
+			queue_writel(queue, TBQP, lower_32_bits(queue->tx_ring_dma));
+		} else {
+			/*
+			 * macb_set_channels(), which is the only way of writing
+			 * to bp->num_queues, is only allowed if
+			 * MACB_CAPS_QUEUE_DISABLE.
+			 */
+			queue_writel(queue, RBQP, MACB_BIT(QUEUE_DISABLE));
+
+			/* Disable all interrupts */
+			queue_writel(queue, IDR, -1);
+			queue_readl(queue, ISR);
+			if (bp->caps & MACB_CAPS_ISR_CLEAR_ON_WRITE)
+				queue_writel(queue, ISR, -1);
+		}
 	}
 }
 
@@ -3900,8 +3915,8 @@ static int gem_set_rxnfc(struct net_device *netdev, struct ethtool_rxnfc *cmd)
 
 	switch (cmd->cmd) {
 	case ETHTOOL_SRXCLSRLINS:
-		if ((cmd->fs.location >= bp->max_tuples)
-				|| (cmd->fs.ring_cookie >= bp->num_queues)) {
+		if (cmd->fs.location >= bp->max_tuples ||
+		    cmd->fs.ring_cookie >= bp->max_num_queues) {
 			ret = -EINVAL;
 			break;
 		}
@@ -3919,6 +3934,54 @@ static int gem_set_rxnfc(struct net_device *netdev, struct ethtool_rxnfc *cmd)
 	return ret;
 }
 
+static void macb_get_channels(struct net_device *netdev,
+			      struct ethtool_channels *ch)
+{
+	struct macb *bp = netdev_priv(netdev);
+
+	ch->max_combined = bp->max_num_queues;
+	ch->combined_count = bp->num_queues;
+}
+
+static int macb_set_channels(struct net_device *netdev,
+			     struct ethtool_channels *ch)
+{
+	struct macb *bp = netdev_priv(netdev);
+	unsigned int old_count = bp->num_queues;
+	unsigned int count = ch->combined_count;
+	bool running = netif_running(bp->dev);
+	int ret = 0;
+
+	if (!(bp->caps & MACB_CAPS_QUEUE_DISABLE))
+		return -EOPNOTSUPP;
+
+	if (!count || ch->rx_count || ch->tx_count)
+		return -EINVAL;
+
+	if (count > bp->max_num_queues)
+		return -EINVAL;
+
+	if (count == old_count)
+		return 0;
+
+	if (running)
+		macb_close(bp->dev);
+
+	bp->num_queues = count;
+	netif_set_real_num_queues(bp->dev, count, count);
+
+	if (running) {
+		ret = macb_open(bp->dev);
+		if (ret) {
+			bp->num_queues = old_count;
+			netif_set_real_num_queues(bp->dev, old_count, old_count);
+			macb_open(bp->dev);
+		}
+	}
+
+	return ret;
+}
+
 static const struct ethtool_ops macb_ethtool_ops = {
 	.get_regs_len		= macb_get_regs_len,
 	.get_regs		= macb_get_regs,
@@ -3934,6 +3997,8 @@ static const struct ethtool_ops macb_ethtool_ops = {
 	.set_link_ksettings     = macb_set_link_ksettings,
 	.get_ringparam		= macb_get_ringparam,
 	.set_ringparam		= macb_set_ringparam,
+	.get_channels		= macb_get_channels,
+	.set_channels		= macb_set_channels,
 };
 
 static const struct ethtool_ops gem_ethtool_ops = {
@@ -3954,10 +4019,12 @@ static const struct ethtool_ops gem_ethtool_ops = {
 	.set_link_ksettings     = macb_set_link_ksettings,
 	.get_ringparam		= macb_get_ringparam,
 	.set_ringparam		= macb_set_ringparam,
-	.get_rxnfc			= gem_get_rxnfc,
-	.set_rxnfc			= gem_set_rxnfc,
-	.get_rx_ring_count		= gem_get_rx_ring_count,
-	.nway_reset			= phy_ethtool_nway_reset,
+	.get_rxnfc		= gem_get_rxnfc,
+	.set_rxnfc		= gem_set_rxnfc,
+	.get_rx_ring_count	= gem_get_rx_ring_count,
+	.nway_reset		= phy_ethtool_nway_reset,
+	.get_channels		= macb_get_channels,
+	.set_channels		= macb_set_channels,
 };
 
 static int macb_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
@@ -4098,9 +4165,9 @@ static int macb_taprio_setup_replace(struct net_device *ndev,
 	size_t i;
 	int err;
 
-	if (conf->num_entries > bp->num_queues) {
+	if (conf->num_entries > bp->max_num_queues) {
 		netdev_err(ndev, "Too many TAPRIO entries: %zu > %d queues\n",
-			   conf->num_entries, bp->num_queues);
+			   conf->num_entries, bp->max_num_queues);
 		return -EINVAL;
 	}
 
@@ -4148,9 +4215,9 @@ static int macb_taprio_setup_replace(struct net_device *ndev,
 
 		/* gate_mask must not select queues outside the valid queues */
 		queue_id = order_base_2(entry->gate_mask);
-		if (queue_id >= bp->num_queues) {
+		if (queue_id >= bp->max_num_queues) {
 			netdev_err(ndev, "Entry %zu: gate_mask 0x%x exceeds queue range (max_queues=%d)\n",
-				   i, entry->gate_mask, bp->num_queues);
+				   i, entry->gate_mask, bp->max_num_queues);
 			err = -EINVAL;
 			goto cleanup;
 		}
@@ -4210,7 +4277,7 @@ static int macb_taprio_setup_replace(struct net_device *ndev,
 	/* All validations passed - proceed with hardware configuration */
 	scoped_guard(spinlock_irqsave, &bp->lock) {
 		/* Disable ENST queues if running before configuring */
-		queue_mask = BIT_U32(bp->num_queues) - 1;
+		queue_mask = BIT_U32(bp->max_num_queues) - 1;
 		gem_writel(bp, ENST_CONTROL,
 			   queue_mask << GEM_ENST_DISABLE_QUEUE_OFFSET);
 
@@ -4245,7 +4312,7 @@ static void macb_taprio_destroy(struct net_device *ndev)
 	unsigned int q;
 
 	netdev_reset_tc(ndev);
-	queue_mask = BIT_U32(bp->num_queues) - 1;
+	queue_mask = BIT_U32(bp->max_num_queues) - 1;
 
 	scoped_guard(spinlock_irqsave, &bp->lock) {
 		/* Single disable command for all queues */
@@ -4253,7 +4320,7 @@ static void macb_taprio_destroy(struct net_device *ndev)
 			   queue_mask << GEM_ENST_DISABLE_QUEUE_OFFSET);
 
 		/* Clear all queue ENST registers in batch */
-		for (q = 0, queue = bp->queues; q < bp->num_queues; ++q, ++queue) {
+		for (q = 0, queue = bp->queues; q < bp->max_num_queues; ++q, ++queue) {
 			queue_writel(queue, ENST_START_TIME, 0);
 			queue_writel(queue, ENST_ON_TIME, 0);
 			queue_writel(queue, ENST_OFF_TIME, 0);
@@ -5512,6 +5579,7 @@ static int macb_probe(struct platform_device *pdev)
 		bp->macb_reg_writel = hw_writel;
 	}
 	bp->num_queues = num_queues;
+	bp->max_num_queues = num_queues;
 	bp->dma_burst_length = macb_config->dma_burst_length;
 	bp->pclk = pclk;
 	bp->hclk = hclk;

-- 
2.53.0

Re: [PATCH net-next 1/2] net: macb: implement ethtool_ops.get|set_channels()

[Jakub Kicinski]

On Thu, 05 Mar 2026 18:20:14 +0100 Théo Lebrun wrote:
> +	if (!(bp->caps & MACB_CAPS_QUEUE_DISABLE))
> +		return -EOPNOTSUPP;

Why not set max to 1 in this case?

> +	if (!count || ch->rx_count || ch->tx_count)
> +		return -EINVAL;

Core should check this for you already

> +	if (count > bp->max_num_queues)
> +		return -EINVAL;

and this

> +	if (count == old_count)
> +		return 0;
> +
> +	if (running)
> +		macb_close(bp->dev);
> +
> +	bp->num_queues = count;
> +	netif_set_real_num_queues(bp->dev, count, count);
> +
> +	if (running) {
> +		ret = macb_open(bp->dev);
> +		if (ret) {
> +			bp->num_queues = old_count;
> +			netif_set_real_num_queues(bp->dev, old_count, old_count);
> +			macb_open(bp->dev);

both macb_open() calls may fail under memory pressure
For new functionality we ask drivers to allocate all necessary
resources upfront then just swap them in and reconfigure HW
-- 
pw-bot: cr

Re: [PATCH net-next 1/2] net: macb: implement ethtool_ops.get|set_channels()

[Théo Lebrun]

On Sat Mar 7, 2026 at 4:09 AM CET, Jakub Kicinski wrote:
> On Thu, 05 Mar 2026 18:20:14 +0100 Théo Lebrun wrote:
>> +	if (!(bp->caps & MACB_CAPS_QUEUE_DISABLE))
>> +		return -EOPNOTSUPP;
>
> Why not set max to 1 in this case?

With !QUEUE_DISABLE, we only know how to run with all queues enabled.
It doesn't imply that max_num_queues == 1.

MACB_CAPS_QUEUE_DISABLE means that the field QUEUE_DISABLE (BIT0) in the
per-queue register RBQP disables queue Rx. If we don't have that
capability we can have multiple queues (if HW supports it) but we must
always run with all enabled.

The correct way to deal with `!(bp->caps & MACB_CAPS_QUEUE_DISABLE)`
would be something like:

static void macb_get_channels(struct net_device *netdev,
			      struct ethtool_channels *ch)
{
	struct macb *bp = netdev_priv(netdev);

	ch->max_combined = bp->max_num_queues;
	ch->combined_count = bp->num_queues;

	if (bp->caps & MACB_CAPS_QUEUE_DISABLE)
		/* we know how disable individual queues */
		ch->min_combined = 1;
	else
		/* we only support running with all queues active */
		ch->min_combined = bp->max_num_queues;
}

But ch->min_combined does not exist.

>
>> +	if (!count || ch->rx_count || ch->tx_count)
>> +		return -EINVAL;
>
> Core should check this for you already
>
>> +	if (count > bp->max_num_queues)
>> +		return -EINVAL;
>
> and this

Noted thanks!

>> +	if (count == old_count)
>> +		return 0;
>> +
>> +	if (running)
>> +		macb_close(bp->dev);
>> +
>> +	bp->num_queues = count;
>> +	netif_set_real_num_queues(bp->dev, count, count);
>> +
>> +	if (running) {
>> +		ret = macb_open(bp->dev);
>> +		if (ret) {
>> +			bp->num_queues = old_count;
>> +			netif_set_real_num_queues(bp->dev, old_count, old_count);
>> +			macb_open(bp->dev);
>
> both macb_open() calls may fail under memory pressure
> For new functionality we ask drivers to allocate all necessary
> resources upfront then just swap them in and reconfigure HW

The main reason we want to set queue count is memory savings. If we take
the Mobileye EyeQ5 SoC, it has a small 32MiB RAM alias usable for DMA.
If we waste it on networking we have less available for the remaining
peripherals. Is there some way we could avoid upfront allocations?

.set_ringparam() can help to avoid memory waste by using many small
queues. But our main target config is a single large queue (common with
AF_XDP zero-copy when userspace wants a single socket). In that case we
waste `(max_num_queues - 1) / max_num_queues` percent: 75% with
max_num_queues=4 on Mobileye EyeQ5 & EyeQ6H, ie Mobileye boards on my
desk at the moment.

I wonder if we'll see GEM IPs that have >4 queues. HW manual indicates
up to 16 is supported.

Thanks Jakub,

--
Théo Lebrun, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com

Re: [PATCH net-next 1/2] net: macb: implement ethtool_ops.get|set_channels()

[Jakub Kicinski]

On Mon, 09 Mar 2026 18:04:06 +0100 Théo Lebrun wrote:
> On Sat Mar 7, 2026 at 4:09 AM CET, Jakub Kicinski wrote:
> > On Thu, 05 Mar 2026 18:20:14 +0100 Théo Lebrun wrote:  
> >> +	if (!(bp->caps & MACB_CAPS_QUEUE_DISABLE))
> >> +		return -EOPNOTSUPP;  
> >
> > Why not set max to 1 in this case?  
> 
> With !QUEUE_DISABLE, we only know how to run with all queues enabled.
> It doesn't imply that max_num_queues == 1.
> 
> MACB_CAPS_QUEUE_DISABLE means that the field QUEUE_DISABLE (BIT0) in the
> per-queue register RBQP disables queue Rx. If we don't have that
> capability we can have multiple queues (if HW supports it) but we must
> always run with all enabled.

Oh, I see! Perhaps just a comment over the check to inform the reader
that the lack of capabilities means all rx queues must be enabled.

> >> +	if (running) {
> >> +		ret = macb_open(bp->dev);
> >> +		if (ret) {
> >> +			bp->num_queues = old_count;
> >> +			netif_set_real_num_queues(bp->dev, old_count, old_count);
> >> +			macb_open(bp->dev);  
> >
> > both macb_open() calls may fail under memory pressure
> > For new functionality we ask drivers to allocate all necessary
> > resources upfront then just swap them in and reconfigure HW  
> 
> The main reason we want to set queue count is memory savings. If we take
> the Mobileye EyeQ5 SoC, it has a small 32MiB RAM alias usable for DMA.
> If we waste it on networking we have less available for the remaining
> peripherals. Is there some way we could avoid upfront allocations?

We've been asking everyone to follow the "pre-allocate resources
paradigm" for a few years now. It has huge benefits for system
reliability. If you don't want to complicate code at this stage
you can support configuring queue count only when the device is down.
But the ask will keep coming back, any time you try to do the close+open

Re: [PATCH net-next 1/2] net: macb: implement ethtool_ops.get|set_channels()

[Théo Lebrun]

On Mon Mar 9, 2026 at 10:15 PM CET, Jakub Kicinski wrote:
> On Mon, 09 Mar 2026 18:04:06 +0100 Théo Lebrun wrote:
>> On Sat Mar 7, 2026 at 4:09 AM CET, Jakub Kicinski wrote:
>> > On Thu, 05 Mar 2026 18:20:14 +0100 Théo Lebrun wrote:  
>> >> +	if (!(bp->caps & MACB_CAPS_QUEUE_DISABLE))
>> >> +		return -EOPNOTSUPP;  
>> >
>> > Why not set max to 1 in this case?  
>> 
>> With !QUEUE_DISABLE, we only know how to run with all queues enabled.
>> It doesn't imply that max_num_queues == 1.
>> 
>> MACB_CAPS_QUEUE_DISABLE means that the field QUEUE_DISABLE (BIT0) in the
>> per-queue register RBQP disables queue Rx. If we don't have that
>> capability we can have multiple queues (if HW supports it) but we must
>> always run with all enabled.
>
> Oh, I see! Perhaps just a comment over the check to inform the reader
> that the lack of capabilities means all rx queues must be enabled.

ACK, will insert a comment.

>> >> +	if (running) {
>> >> +		ret = macb_open(bp->dev);
>> >> +		if (ret) {
>> >> +			bp->num_queues = old_count;
>> >> +			netif_set_real_num_queues(bp->dev, old_count, old_count);
>> >> +			macb_open(bp->dev);  
>> >
>> > both macb_open() calls may fail under memory pressure
>> > For new functionality we ask drivers to allocate all necessary
>> > resources upfront then just swap them in and reconfigure HW  
>> 
>> The main reason we want to set queue count is memory savings. If we take
>> the Mobileye EyeQ5 SoC, it has a small 32MiB RAM alias usable for DMA.
>> If we waste it on networking we have less available for the remaining
>> peripherals. Is there some way we could avoid upfront allocations?
>
> We've been asking everyone to follow the "pre-allocate resources
> paradigm" for a few years now. It has huge benefits for system
> reliability. If you don't want to complicate code at this stage
> you can support configuring queue count only when the device is down.
> But the ask will keep coming back, any time you try to do the close+open

I'll iterate with a .set_channels() implementation that returns -EBUSY
if netif_running(), making sure to leave a comment explaining the
tradeoff.

Thanks Jakub,

--
Théo Lebrun, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com

[PATCH net-next 2/2] net: macb: distribute evenly Tx SRAM segments

[Théo Lebrun]

GEM has registers to configure the Tx SRAM segments distribution across
queues. The reset value is apprioriate (even spread) but we need to
care if/when number of active queues is modified (or if we inherited
unevenly initialised hardware from bootloader).

To distribute segments, we take as input the number of queues
(bp->num_queues) and the number of segments (found inside DCFG6).
Its output is a number of segments for each queue, formatted as
powers-of-two (eg 2 for queue 0 means it has 2^2=4 segments).

As the distribution logic is quite complex (at least its initial
versions had bugs), it is kunit-tested and those tests live at the end
of macb_main.c. To test:

⟩ env --unset=CROSS_COMPILE make ARCH=um mrproper
⟩ env --unset=CROSS_COMPILE ./tools/testing/kunit/kunit.py run \
        --kconfig_add CONFIG_NET=y \
        --kconfig_add CONFIG_COMMON_CLK=y \
        --kconfig_add CONFIG_MACB=y 'macb*'

Signed-off-by: Théo Lebrun <theo.lebrun@bootlin.com>
---
 drivers/net/ethernet/cadence/Kconfig     |   6 ++
 drivers/net/ethernet/cadence/macb.h      |   5 ++
 drivers/net/ethernet/cadence/macb_main.c | 135 +++++++++++++++++++++++++++++++
 3 files changed, 146 insertions(+)

diff --git a/drivers/net/ethernet/cadence/Kconfig b/drivers/net/ethernet/cadence/Kconfig
index 5b2a461dfd28..3ae7123352f5 100644
--- a/drivers/net/ethernet/cadence/Kconfig
+++ b/drivers/net/ethernet/cadence/Kconfig
@@ -51,4 +51,10 @@ config MACB_PCI
 	  To compile this driver as a module, choose M here: the module
 	  will be called macb_pci.
 
+config MACB_KUNIT_TEST
+	bool "KUnit test for MACB" if !KUNIT_ALL_TESTS
+	depends on KUNIT
+	depends on MACB
+	default KUNIT_ALL_TESTS
+
 endif # NET_VENDOR_CADENCE
diff --git a/drivers/net/ethernet/cadence/macb.h b/drivers/net/ethernet/cadence/macb.h
index 30fa65e2bdf2..81fdd17b34db 100644
--- a/drivers/net/ethernet/cadence/macb.h
+++ b/drivers/net/ethernet/cadence/macb.h
@@ -193,6 +193,9 @@
 #define GEM_TXBDCTRL	0x04cc /* TX Buffer Descriptor control register */
 #define GEM_RXBDCTRL	0x04d0 /* RX Buffer Descriptor control register */
 
+#define GEM_TXQSEGALLOC_LOWER	0x05A0 /* Tx queue segment allocation (low) */
+#define GEM_TXQSEGALLOC_UPPER	0x05A4 /* Tx queue segment allocation (high) */
+
 /* Screener Type 2 match registers */
 #define GEM_SCRT2		0x540
 
@@ -543,6 +546,8 @@
 #define GEM_PBUF_CUTTHRU_SIZE			1
 #define GEM_DAW64_OFFSET			23
 #define GEM_DAW64_SIZE				1
+#define GEM_SEGMENTS_BIT_SIZE_OFFSET		16
+#define GEM_SEGMENTS_BIT_SIZE_SIZE		3
 
 /* Bitfields in DCFG8. */
 #define GEM_T1SCR_OFFSET			24
diff --git a/drivers/net/ethernet/cadence/macb_main.c b/drivers/net/ethernet/cadence/macb_main.c
index bac83a2b4c4d..022577756eab 100644
--- a/drivers/net/ethernet/cadence/macb_main.c
+++ b/drivers/net/ethernet/cadence/macb_main.c
@@ -2751,6 +2751,59 @@ static u32 macb_dbw(struct macb *bp)
 	}
 }
 
+/*
+ * Distribute evenly available segments across queues. The computation is
+ * complex because (1) segments are counted in powers of two and (2) a queue
+ * can only use up to 8 segments. There are four types of cases:
+ *  - Sharing all segments equally is doable. Take num_queues=4 and
+ *    num_segments=16. Each queue will get 2^2=4 segments.
+ *  - Sharing all segments is doable. Take num_queues=5 and num_segments=16.
+ *    Three queues will get 2^2=4 segments and two will get 2^1=2 segments.
+ *  - Sharing all segments is not doable because not enough queues are
+ *    available. Take num_queues=1 and num_segments=16; queue 0 can only have 8
+ *    segments.
+ *  - Sharing all segments is not doable because not enough segments are
+ *    available. Take num_queues=4 and num_segments=2.
+ *
+ * We start by computing the power each queue will have. For num_queues=5 and
+ * num_segments=16, each queue will have at least 2^1 segments. That leaves us
+ * with remaining_segments=6. If we increase the power for a queue, we get a
+ * delta of 2 (2^2-2^1). The first three queues will therefore be advantaged
+ * and each have 2^2 segments. The remaining 2 queues will only have 2^1
+ * segments.
+ */
+static u64 gem_sram_distribute_segments(unsigned int num_queues,
+					unsigned int num_segments)
+{
+	unsigned int pow, remaining_segments, i;
+	unsigned int num_advantaged_queues = 0;
+	u64 val = 0;
+
+	/* pow=0 for all queues. ilog2(0) is dangerous. */
+	if (num_queues >= num_segments)
+		return 0;
+
+	pow = min(ilog2(num_segments / num_queues), 3);
+	remaining_segments = num_segments - num_queues * (1U << pow);
+
+	/*
+	 * We can only distribute remaining segments if (1) there are remaining
+	 * segments and (2) we did not reach the max segments per queue (2^3).
+	 */
+	if (remaining_segments != 0 && pow != 3) {
+		unsigned int delta = (1U << (pow + 1)) - (1U << pow);
+
+		num_advantaged_queues = remaining_segments / delta;
+	}
+
+	for (i = 0; i < num_advantaged_queues; i++)
+		val |= ((pow + 1) & 0b11) << (i * 4);
+	for (i = num_advantaged_queues; i < num_queues; i++)
+		val |= (pow & 0b11) << (i * 4);
+
+	return val;
+}
+
 /* Configure the receive DMA engine
  * - use the correct receive buffer size
  * - set best burst length for DMA operations
@@ -2832,6 +2885,19 @@ static void macb_init_hw(struct macb *bp)
 	if (bp->caps & MACB_CAPS_JUMBO)
 		bp->rx_frm_len_mask = MACB_RX_JFRMLEN_MASK;
 
+	/*
+	 * Distribute Tx SRAM segments evenly based on active number of queues.
+	 */
+	if (macb_is_gem(bp)) {
+		unsigned int num_segments;
+		u64 val;
+
+		num_segments = 1U << GEM_BFEXT(SEGMENTS_BIT_SIZE, gem_readl(bp, DCFG6));
+		val = gem_sram_distribute_segments(bp->num_queues, num_segments);
+		gem_writel(bp, TXQSEGALLOC_LOWER, val);
+		gem_writel(bp, TXQSEGALLOC_UPPER, val >> 32);
+	}
+
 	macb_configure_dma(bp);
 
 	/* Enable RX partial store and forward and set watermark */
@@ -6031,3 +6097,72 @@ MODULE_LICENSE("GPL");
 MODULE_DESCRIPTION("Cadence MACB/GEM Ethernet driver");
 MODULE_AUTHOR("Haavard Skinnemoen (Atmel)");
 MODULE_ALIAS("platform:macb");
+
+#ifdef CONFIG_MACB_KUNIT_TEST
+#include <kunit/test.h>
+
+struct macb_sram_segments_case {
+	unsigned int num_queues, num_segments;
+};
+
+static void macb_sram_segments_test(struct kunit *test)
+{
+	const struct macb_sram_segments_case *p = test->param_value;
+	u64 val = gem_sram_distribute_segments(p->num_queues, p->num_segments);
+	unsigned int i, sum_segments = 0, max_assigned_segments;
+	unsigned int num_queues = min(p->num_queues, p->num_segments);
+
+	for (i = 0; i < num_queues; i++) {
+		unsigned int q_segments = (val >> (i * 4)) & 0b11;
+
+		q_segments = 1U << q_segments;
+		sum_segments += q_segments;
+		KUNIT_ASSERT_GT_MSG(test, q_segments, 0, "queue %d, val %#llx", i, val);
+	}
+
+	for (i = num_queues; i < 16; i++) {
+		unsigned int pow = (val >> (i * 4)) & 0b11;
+
+		KUNIT_ASSERT_EQ_MSG(test, pow, 0, "queue %d, val %#llx", i, val);
+	}
+
+	max_assigned_segments = min(p->num_segments, 8 * p->num_queues);
+	KUNIT_ASSERT_EQ_MSG(test, sum_segments, max_assigned_segments, "val %#llx", val);
+}
+
+struct macb_sram_segments_case macb_sram_segments_cases[] = {
+	/* num_segments can only be powers of two. */
+	{ .num_queues = 4,  .num_segments = 2 },
+	{ .num_queues = 1,  .num_segments = 16 },
+	{ .num_queues = 4,  .num_segments = 16 },
+	{ .num_queues = 5,  .num_segments = 16 },
+	{ .num_queues = 15, .num_segments = 16 },
+	{ .num_queues = 16, .num_segments = 16 },
+};
+
+static void macb_sram_segments_case_desc(struct macb_sram_segments_case *t, char *desc)
+{
+	u64 val = gem_sram_distribute_segments(t->num_queues, t->num_segments);
+
+	snprintf(desc, KUNIT_PARAM_DESC_SIZE,
+		 "num_queues=%d num_segments=%d TXQSEGALLOC=%#llx",
+		 t->num_queues, t->num_segments, val);
+}
+
+KUNIT_ARRAY_PARAM(macb_sram_segments,
+		  macb_sram_segments_cases,
+		  macb_sram_segments_case_desc);
+
+static struct kunit_case macb_test_cases[] = {
+	KUNIT_CASE_PARAM(macb_sram_segments_test, macb_sram_segments_gen_params),
+	{}
+};
+
+static struct kunit_suite macb_test_suite = {
+	.name = "macb",
+	.test_cases = macb_test_cases,
+};
+
+kunit_test_suite(macb_test_suite);
+
+#endif

-- 
2.53.0

Re: [PATCH net-next 2/2] net: macb: distribute evenly Tx SRAM segments

[Jakub Kicinski]

On Thu, 05 Mar 2026 18:20:15 +0100 Théo Lebrun wrote:
>  MODULE_DESCRIPTION("Cadence MACB/GEM Ethernet driver");
>  MODULE_AUTHOR("Haavard Skinnemoen (Atmel)");
>  MODULE_ALIAS("platform:macb");

these macros should be at the end of the file
you should probably move the test to a new file

> +#ifdef CONFIG_MACB_KUNIT_TEST
> +#include <kunit/test.h>
> +
> +struct macb_sram_segments_case {
> +	unsigned int num_queues, num_segments;
> +};