Skip to contentSkip to conversation

net: ks8851: Reinstate disabling of BHs around IRQ handler

PATCH Superseded Marek Vasut 3 participants 13 messages Apr 14 – Apr 18

Subsystems (3)

v3 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

PATCH v3 1/2 net: ks8851: Reinstate disabling of BHs around IRQ handler

If CONFIG_PREEMPT_RT=y is set AND the driver executes ks8851_irq() AND KSZ_ISR register bit IRQ_RXI is set AND ks8851_rx_pkts() detects that there are packets in the RX FIFO, then netdev_alloc_skb_ip_align() is called to allocate SKBs. If netdev_alloc_skb_ip_align() is called with BH enabled, local_bh_enable() at the end of netdev_alloc_skb_ip_align() will call __local_bh_enable_ip(), which will call __do_softirq(), which may trigger net_tx_action() softirq, which may ultimately call the xmit callback ks8851_start_xmit_par(). The ks8851_start_xmit_par() will try to lock struct ks8851_net_par .lock spinlock, which is already locked by ks8851_irq() from which ks8851_start_xmit_par() was called. This leads to a deadlock, which is reported by the kernel, including a trace listed below.

Fix the problem by disabling BH around critical sections, including the IRQ handler, thus preventing the net_tx_action() softirq from triggering during these critical sections. The net_tx_action() softirq is triggered at the end of the IRQ handler, once all the other IRQ handler actions have been completed.

__schedule from schedule_rtlock+0x1c/0x34 schedule_rtlock from rtlock_slowlock_locked+0x548/0x904 rtlock_slowlock_locked from rt_spin_lock+0x60/0x9c rt_spin_lock from ks8851_start_xmit_par+0x74/0x1a8 ks8851_start_xmit_par from netdev_start_xmit+0x20/0x44 netdev_start_xmit from dev_hard_start_xmit+0xd0/0x188 dev_hard_start_xmit from sch_direct_xmit+0xb8/0x25c sch_direct_xmit from __qdisc_run+0x1f8/0x4ec __qdisc_run from qdisc_run+0x1c/0x28
qdisc_run from net_tx_action+0x1f0/0x268 net_tx_action from handle_softirqs+0x1a4/0x270 handle_softirqs from __local_bh_enable_ip+0xcc/0xe0 __local_bh_enable_ip from __alloc_skb+0xd8/0x128 __alloc_skb from __netdev_alloc_skb+0x3c/0x19c __netdev_alloc_skb from ks8851_irq+0x388/0x4d4 ks8851_irq from irq_thread_fn+0x24/0x64 irq_thread_fn from irq_thread+0x178/0x28c irq_thread from kthread+0x12c/0x138
kthread from ret_from_fork+0x14/0x28

Fixes e0863634bf9f ("net: ks8851: Queue RX packets in IRQ handler instead of disabling BHs")

---
Cc: "David S. Miller" <[email protected]>
Cc: Andrew Lunn <[email protected]>
Cc: Eric Dumazet <[email protected]>
Cc: Jakub Kicinski <[email protected]>
Cc: Nicolai Buchwitz <[email protected]>
Cc: Paolo Abeni <[email protected]>
Cc: Ronald Wahl <[email protected]>
Cc: Sebastian Andrzej Siewior <[email protected]>
Cc: Yicong Hui <[email protected]>
Cc: [email protected]
Cc: [email protected]
---
V2: Register dedicated IRQ handler wrapper which disables BH for the
parallel variant of the MAC, the variant which uses spinlocks as the locking primitive. Use stock IRQ handler with BH unchanged for the SPI variant, which uses mutexes as locking primitive.
V3: Switch to spin_lock_bh(), update commit message

---
 drivers/net/ethernet/micrel/ks8851.h        |  6 +-
 drivers/net/ethernet/micrel/ks8851_common.c | 64 +++++++++------------
 drivers/net/ethernet/micrel/ks8851_par.c    | 15 ++---
 drivers/net/ethernet/micrel/ks8851_spi.c    | 11 ++--
 4 files changed, 38 insertions(+), 58 deletions(-)
drivers/net/ethernet/micrel/ks8851.h +2 -4
index 31f75b4a67fd7..b795a3a605711 100644
@@ -408,10 +408,8 @@ struct ks8851_net {
 	struct gpio_desc	*gpio;
 	struct mii_bus		*mii_bus;
 
-	void			(*lock)(struct ks8851_net *ks,
-					unsigned long *flags);
-	void			(*unlock)(struct ks8851_net *ks,
-					  unsigned long *flags);
+	void			(*lock)(struct ks8851_net *ks);
+	void			(*unlock)(struct ks8851_net *ks);
 	unsigned int		(*rdreg16)(struct ks8851_net *ks,
 					   unsigned int reg);
 	void			(*wrreg16)(struct ks8851_net *ks,
drivers/net/ethernet/micrel/ks8851_common.c +26 -38

243 lines not on this page · raw message

drivers/net/ethernet/micrel/ks8851_par.c +6 -9
index 78695be2570bf..9f1c33f6ddec0 100644
@@ -55,29 +55,27 @@ struct ks8851_net_par {
 /**
  * ks8851_lock_par - register access lock
  * @ks: The chip state
- * @flags: Spinlock flags
  *
  * Claim chip register access lock
  */
-static void ks8851_lock_par(struct ks8851_net *ks, unsigned long *flags)
+static void ks8851_lock_par(struct ks8851_net *ks)
 {
 	struct ks8851_net_par *ksp = to_ks8851_par(ks);
 
-	spin_lock_irqsave(&ksp->lock, *flags);
+	spin_lock_bh(&ksp->lock);
 }
 
 /**
  * ks8851_unlock_par - register access unlock
  * @ks: The chip state
- * @flags: Spinlock flags
  *
  * Release chip register access lock
  */
-static void ks8851_unlock_par(struct ks8851_net *ks, unsigned long *flags)
+static void ks8851_unlock_par(struct ks8851_net *ks)
 {
 	struct ks8851_net_par *ksp = to_ks8851_par(ks);
 
-	spin_unlock_irqrestore(&ksp->lock, *flags);
+	spin_unlock_bh(&ksp->lock);
 }
 
 /**
@@ -233,7 +231,6 @@ static netdev_tx_t ks8851_start_xmit_par(struct sk_buff *skb,
 {
 	struct ks8851_net *ks = netdev_priv(dev);
 	netdev_tx_t ret = NETDEV_TX_OK;
-	unsigned long flags;
 	unsigned int txqcr;
 	u16 txmir;
 	int err;
@@ -241,7 +238,7 @@ static netdev_tx_t ks8851_start_xmit_par(struct sk_buff *skb,
 	netif_dbg(ks, tx_queued, ks->netdev,
 		  "%s: skb %p, %d@%p\n", __func__, skb, skb->len, skb->data);
 
-	ks8851_lock_par(ks, &flags);
+	ks8851_lock_par(ks);
 
 	txmir = ks8851_rdreg16_par(ks, KS_TXMIR) & 0x1fff;
 
@@ -262,7 +259,7 @@ static netdev_tx_t ks8851_start_xmit_par(struct sk_buff *skb,
 		ret = NETDEV_TX_BUSY;
 	}
 
-	ks8851_unlock_par(ks, &flags);
+	ks8851_unlock_par(ks);
 
 	return ret;
 }
drivers/net/ethernet/micrel/ks8851_spi.c +4 -7
index a161ae45743ab..b9e68520278d0 100644
@@ -71,11 +71,10 @@ struct ks8851_net_spi {
 /**
  * ks8851_lock_spi - register access lock
  * @ks: The chip state
- * @flags: Spinlock flags
  *
  * Claim chip register access lock
  */
-static void ks8851_lock_spi(struct ks8851_net *ks, unsigned long *flags)
+static void ks8851_lock_spi(struct ks8851_net *ks)
 {
 	struct ks8851_net_spi *kss = to_ks8851_spi(ks);
 
@@ -85,11 +84,10 @@ static void ks8851_lock_spi(struct ks8851_net *ks, unsigned long *flags)
 /**
  * ks8851_unlock_spi - register access unlock
  * @ks: The chip state
- * @flags: Spinlock flags
  *
  * Release chip register access lock
  */
-static void ks8851_unlock_spi(struct ks8851_net *ks, unsigned long *flags)
+static void ks8851_unlock_spi(struct ks8851_net *ks)
 {
 	struct ks8851_net_spi *kss = to_ks8851_spi(ks);
 
@@ -309,7 +307,6 @@ static void ks8851_tx_work(struct work_struct *work)
 	struct ks8851_net_spi *kss;
 	unsigned short tx_space;
 	struct ks8851_net *ks;
-	unsigned long flags;
 	struct sk_buff *txb;
 	bool last;
 
@@ -317,7 +314,7 @@ static void ks8851_tx_work(struct work_struct *work)
 	ks = &kss->ks8851;
 	last = skb_queue_empty(&ks->txq);
 
-	ks8851_lock_spi(ks, &flags);
+	ks8851_lock_spi(ks);
 
 	while (!last) {
 		txb = skb_dequeue(&ks->txq);
@@ -343,7 +340,7 @@ static void ks8851_tx_work(struct work_struct *work)
 	ks->tx_space = tx_space;
 	spin_unlock_bh(&ks->statelock);
 
-	ks8851_unlock_spi(ks, &flags);
+	ks8851_unlock_spi(ks);
 }
 
 /**
Signature · 1 line
2.53.0
#1 lore raw

PATCH v3 2/2 net: ks8851: Avoid excess softirq scheduling

The code injects a packet into netif_rx() repeatedly, which will add it to its internal NAPI and schedule a softirq, and process it. It is more efficient to queue multiple packets and process them all at the local_bh_enable() time.

Fixes e0863634bf9f ("net: ks8851: Queue RX packets in IRQ handler instead of disabling BHs")

---
Cc: "David S. Miller" <[email protected]>
Cc: Andrew Lunn <[email protected]>
Cc: Eric Dumazet <[email protected]>
Cc: Jakub Kicinski <[email protected]>
Cc: Nicolai Buchwitz <[email protected]>
Cc: Paolo Abeni <[email protected]>
Cc: Ronald Wahl <[email protected]>
Cc: Sebastian Andrzej Siewior <[email protected]>
Cc: Yicong Hui <[email protected]>
Cc: [email protected]
Cc: [email protected]
---
V3: New patch

---
 drivers/net/ethernet/micrel/ks8851_common.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
drivers/net/ethernet/micrel/ks8851_common.c +4 -1
index 6c375647b24de..4afbb40bc0e4a 100644
@@ -373,9 +373,12 @@ static irqreturn_t ks8851_irq(int irq, void *_ks)
 	if (status & IRQ_LCI)
 		mii_check_link(&ks->mii);
 
-	if (status & IRQ_RXI)
+	if (status & IRQ_RXI) {
+		local_bh_disable();
 		while ((skb = __skb_dequeue(&rxq)))
 			netif_rx(skb);
+		local_bh_enable();
+	}
 
 	return IRQ_HANDLED;
 }
Signature · 1 line
2.53.0
Reviewed-by on v3 2/2
#2 lore raw

Re: [net,PATCH v3 2/2] net: ks8851: Avoid excess softirq scheduling

Quoted 8 lines · Marek Vasut

The code injects a packet into netif_rx() repeatedly, which will add it to its internal NAPI and schedule a softirq, and process it. It is more efficient to queue multiple packets and process them all at the local_bh_enable() time.

Fixes e0863634bf9f ("net: ks8851: Queue RX packets in IRQ handler instead of disabling BHs")

Sebastian

Reviewed-by on v3 1/2
#3 lore raw
In reply to Marek Vasut
Quoted 4 lines · Marek Vasut

If CONFIG_PREEMPT_RT=y is set AND the driver executes ks8851_irq() AND KSZ_ISR register bit IRQ_RXI is set AND ks8851_rx_pkts() detects that there are packets in the RX FIFO, then netdev_alloc_skb_ip_align() is called to allocate SKBs. If netdev_alloc_skb_ip_align() is called with

Quoted 8 lines

BH enabled, local_bh_enable() at the end of netdev_alloc_skb_ip_align() will call __local_bh_enable_ip(), which will call __do_softirq(), which may trigger net_tx_action() softirq, which may ultimately call the xmit callback ks8851_start_xmit_par(). The ks8851_start_xmit_par() will try to lock struct ks8851_net_par .lock spinlock, which is already locked by ks8851_irq() from which ks8851_start_xmit_par() was called. This leads to a deadlock, which is reported by the kernel, including a trace listed below.

#1 [received RX packet and a] TX packet has been sent
#2 Driver enables TX queue via netif_wake_queue() which schedules TX
softirq to queue packets for this device.
#2 After spin_unlock_bh(&ks->statelock) the pending softirqs will be
processed
#3 This deadlocks because of recursive locking via ks8851_net::lock in
ks8851_irq() and ks8851_start_xmit_par().

This is what happens since commit 0913ec336a6c0 ("net: ks8851: Fix deadlock with the SPI chip variant"). Before that commit the softirq execution will be picked up by netdev_alloc_skb_ip_align() and requires PREEMPT_RT and a RX packet in #1 to trigger the deadlock.

Quoted 17 lines

Fix the problem by disabling BH around critical sections, including the IRQ handler, thus preventing the net_tx_action() softirq from triggering during these critical sections. The net_tx_action() softirq is triggered at the end of the IRQ handler, once all the other IRQ handler actions have been completed.

__schedule from schedule_rtlock+0x1c/0x34 schedule_rtlock from rtlock_slowlock_locked+0x548/0x904 rtlock_slowlock_locked from rt_spin_lock+0x60/0x9c rt_spin_lock from ks8851_start_xmit_par+0x74/0x1a8 ks8851_start_xmit_par from netdev_start_xmit+0x20/0x44 netdev_start_xmit from dev_hard_start_xmit+0xd0/0x188 dev_hard_start_xmit from sch_direct_xmit+0xb8/0x25c sch_direct_xmit from __qdisc_run+0x1f8/0x4ec __qdisc_run from qdisc_run+0x1c/0x28
qdisc_run from net_tx_action+0x1f0/0x268 net_tx_action from handle_softirqs+0x1a4/0x270

Quoted 8 lines

handle_softirqs from __local_bh_enable_ip+0xcc/0xe0 __local_bh_enable_ip from __alloc_skb+0xd8/0x128 __alloc_skb from __netdev_alloc_skb+0x3c/0x19c __netdev_alloc_skb from ks8851_irq+0x388/0x4d4 ks8851_irq from irq_thread_fn+0x24/0x64 irq_thread_fn from irq_thread+0x178/0x28c irq_thread from kthread+0x12c/0x138
kthread from ret_from_fork+0x14/0x28

The backtrace here and the description is based on an older kernel. However

Quoted 3 lines
Fixes e0863634bf9f ("net: ks8851: Queue RX packets in IRQ handler instead of disabling BHs")

Sebastian

#4 lore raw
Quoted 20 lines · Sebastian Andrzej Siewior
Quoted 4 lines · Marek Vasut

If CONFIG_PREEMPT_RT=y is set AND the driver executes ks8851_irq() AND KSZ_ISR register bit IRQ_RXI is set AND ks8851_rx_pkts() detects that there are packets in the RX FIFO, then netdev_alloc_skb_ip_align() is called to allocate SKBs. If netdev_alloc_skb_ip_align() is called with

Quoted 8 lines

BH enabled, local_bh_enable() at the end of netdev_alloc_skb_ip_align() will call __local_bh_enable_ip(), which will call __do_softirq(), which may trigger net_tx_action() softirq, which may ultimately call the xmit callback ks8851_start_xmit_par(). The ks8851_start_xmit_par() will try to lock struct ks8851_net_par .lock spinlock, which is already locked by ks8851_irq() from which ks8851_start_xmit_par() was called. This leads to a deadlock, which is reported by the kernel, including a trace listed below.

#1 [received RX packet and a] TX packet has been sent
#2 Driver enables TX queue via netif_wake_queue() which schedules TX
softirq to queue packets for this device.
#2 After spin_unlock_bh(&ks->statelock) the pending softirqs will be
processed
#3 This deadlocks because of recursive locking via ks8851_net::lock in

Quoted 6 lines
    ks8851_irq() and ks8851_start_xmit_par().

This is what happens since commit 0913ec336a6c0 ("net: ks8851: Fix deadlock with the SPI chip variant"). Before that commit the softirq execution will be picked up by netdev_alloc_skb_ip_align() and requires PREEMPT_RT and a RX packet in #1 to trigger the deadlock.

Do you want me to add this into the V4 commit message ?

Quoted 26 lines
Quoted 17 lines

Fix the problem by disabling BH around critical sections, including the IRQ handler, thus preventing the net_tx_action() softirq from triggering during these critical sections. The net_tx_action() softirq is triggered at the end of the IRQ handler, once all the other IRQ handler actions have been completed.

__schedule from schedule_rtlock+0x1c/0x34 schedule_rtlock from rtlock_slowlock_locked+0x548/0x904 rtlock_slowlock_locked from rt_spin_lock+0x60/0x9c rt_spin_lock from ks8851_start_xmit_par+0x74/0x1a8 ks8851_start_xmit_par from netdev_start_xmit+0x20/0x44 netdev_start_xmit from dev_hard_start_xmit+0xd0/0x188 dev_hard_start_xmit from sch_direct_xmit+0xb8/0x25c sch_direct_xmit from __qdisc_run+0x1f8/0x4ec __qdisc_run from qdisc_run+0x1c/0x28
qdisc_run from net_tx_action+0x1f0/0x268 net_tx_action from handle_softirqs+0x1a4/0x270

Quoted 8 lines

handle_softirqs from __local_bh_enable_ip+0xcc/0xe0 __local_bh_enable_ip from __alloc_skb+0xd8/0x128 __alloc_skb from __netdev_alloc_skb+0x3c/0x19c __netdev_alloc_skb from ks8851_irq+0x388/0x4d4 ks8851_irq from irq_thread_fn+0x24/0x64 irq_thread_fn from irq_thread+0x178/0x28c irq_thread from kthread+0x12c/0x138
kthread from ret_from_fork+0x14/0x28

Quoted 2 lines

The backtrace here and the description is based on an older kernel. However

I actually did update the backtrace in V3 with the one from current next 20260413 .

#5 lore raw
Quoted 6 lines · Marek Vasut
Quoted 4 lines

This is what happens since commit 0913ec336a6c0 ("net: ks8851: Fix deadlock with the SPI chip variant"). Before that commit the softirq execution will be picked up by netdev_alloc_skb_ip_align() and requires PREEMPT_RT and a RX packet in #1 to trigger the deadlock.

Do you want me to add this into the V4 commit message ?

The description does not match the code since the commit mentioned above.

Quoted 26 lines
Quoted 26 lines
Quoted 25 lines

Fix the problem by disabling BH around critical sections, including the IRQ handler, thus preventing the net_tx_action() softirq from triggering during these critical sections. The net_tx_action() softirq is triggered at the end of the IRQ handler, once all the other IRQ handler actions have been completed.

__schedule from schedule_rtlock+0x1c/0x34 schedule_rtlock from rtlock_slowlock_locked+0x548/0x904 rtlock_slowlock_locked from rt_spin_lock+0x60/0x9c rt_spin_lock from ks8851_start_xmit_par+0x74/0x1a8 ks8851_start_xmit_par from netdev_start_xmit+0x20/0x44 netdev_start_xmit from dev_hard_start_xmit+0xd0/0x188 dev_hard_start_xmit from sch_direct_xmit+0xb8/0x25c sch_direct_xmit from __qdisc_run+0x1f8/0x4ec __qdisc_run from qdisc_run+0x1c/0x28
qdisc_run from net_tx_action+0x1f0/0x268 net_tx_action from handle_softirqs+0x1a4/0x270 handle_softirqs from __local_bh_enable_ip+0xcc/0xe0 __local_bh_enable_ip from __alloc_skb+0xd8/0x128 __alloc_skb from __netdev_alloc_skb+0x3c/0x19c __netdev_alloc_skb from ks8851_irq+0x388/0x4d4 ks8851_irq from irq_thread_fn+0x24/0x64 irq_thread_fn from irq_thread+0x178/0x28c irq_thread from kthread+0x12c/0x138
kthread from ret_from_fork+0x14/0x28

Quoted 4 lines
Quoted 2 lines

The backtrace here and the description is based on an older kernel. However

I actually did update the backtrace in V3 with the one from current next 20260413 .

That would be from yesterday and the change is merged since v6.10. But why is the softirq starting from __netdev_alloc_skb() instead of spin_unlock_bh(&ks->statelock)? After that unlock, the softirq must be processed and __netdev_alloc_skb() _could_ observe pending softirqs but not from ks8851.

Sebastian

#6 lore raw
Quoted 10 lines · Sebastian Andrzej Siewior
Quoted 6 lines · Marek Vasut
Quoted 4 lines

This is what happens since commit 0913ec336a6c0 ("net: ks8851: Fix deadlock with the SPI chip variant"). Before that commit the softirq execution will be picked up by netdev_alloc_skb_ip_align() and requires PREEMPT_RT and a RX packet in #1 to trigger the deadlock.

Do you want me to add this into the V4 commit message ?

The description does not match the code since the commit mentioned above.

I hope the V4 commit message is a bit better.

Quoted 10 lines
Quoted 4 lines
Quoted 2 lines

The backtrace here and the description is based on an older kernel. However

I actually did update the backtrace in V3 with the one from current next 20260413 .

That would be from yesterday and the change is merged since v6.10. But why is the softirq starting from __netdev_alloc_skb() instead of spin_unlock_bh(&ks->statelock)? After that unlock, the softirq must be processed and __netdev_alloc_skb() _could_ observe pending softirqs but not from ks8851.

Because __netdev_alloc_skb() also enables/disables BH , see the "else" branch:

759 struct sk_buff *__netdev_alloc_skb(struct net_device *dev,
unsigned int len,
760 gfp_t gfp_mask)
761 {
...
786 if (in_hardirq() || irqs_disabled()) {

 787                 nc = this_cpu_ptr(&netdev_alloc_cache);
 788                 data = page_frag_alloc(nc, len, gfp_mask);
 789                 pfmemalloc = page_frag_cache_is_pfmemalloc(nc);

790 } else {
791 local_bh_disable();
792 local_lock_nested_bh(&napi_alloc_cache.bh_lock);
...
798 local_unlock_nested_bh(&napi_alloc_cache.bh_lock);
799 local_bh_enable();
...

#7 lore raw
Quoted 4 lines · Marek Vasut
Quoted 3 lines

spin_unlock_bh(&ks->statelock)? After that unlock, the softirq must be processed and __netdev_alloc_skb() _could_ observe pending softirqs but not from ks8851.

Because __netdev_alloc_skb() also enables/disables BH , see the "else"

Yes. But there is no softirq raised in that part. That softirq is raised by netif_wake_queue() within a bh disabled section. Therefore upon the unlock the softirq must be invoked.
After that, rhe allocation later on may invoke softirqs which were raised but I don't see how ks8851 can be part of it.
Before commit 0913ec336a6c0 ("net: ks8851: Fix deadlock with the SPI chip variant") there was no _bh around it meaning the softirq was raised but not invoked immediately. This happened on the bh unlock during memory allocation. Therefore I am saying this backtrace is from an older kernel.

If there is a flaw in my the theory please explain _how_ you managed that get that backtrace. I am sure it must have from an older kernel and _now_ this lockup also happens on !RT kernels (except for the SPI platform).

Sebastian

#8 lore raw
Quoted 6 lines · Sebastian Andrzej Siewior
Quoted 4 lines · Marek Vasut
Quoted 3 lines

spin_unlock_bh(&ks->statelock)? After that unlock, the softirq must be processed and __netdev_alloc_skb() _could_ observe pending softirqs but not from ks8851.

Because __netdev_alloc_skb() also enables/disables BH , see the "else"

Quoted 10 lines

Yes. But there is no softirq raised in that part. That softirq is raised by netif_wake_queue() within a bh disabled section. Therefore upon the unlock the softirq must be invoked.
After that, rhe allocation later on may invoke softirqs which were raised but I don't see how ks8851 can be part of it.
Before commit 0913ec336a6c0 ("net: ks8851: Fix deadlock with the SPI chip variant") there was no _bh around it meaning the softirq was raised but not invoked immediately. This happened on the bh unlock during memory allocation. Therefore I am saying this backtrace is from an older kernel.

I actually did update the backtrace in V3 with the one from next 20260413 that contained b44596ffe1b4 ("ARM: Allow to enable RT") from stable-rt/v6.12-rt-rebase branch [1] .

I think I misunderstood the usage of "softirq is raised" vs. "softirq is invoked" above . Is it possible that there was an already raised softirq before the threaded IRQ handler was invoked, and __netdev_alloc_skb() is what invoked that softirq ?

Quoted 4 lines

If there is a flaw in my the theory please explain _how_ you managed that get that backtrace. I am sure it must have from an older kernel and _now_ this lockup also happens on !RT kernels (except for the SPI platform).

I used [1] , with PREEMPT_RT enabled , on stm32mp157c SoC . I ran iperf3 -s on the stm32 side, iperf3 -c 192.168.1.2 -t 0 --bidir on the hostpc side. The backtrace happened shortly after.

#9 lore raw
Quoted 11 lines · Marek Vasut
Quoted 2 lines

memory allocation. Therefore I am saying this backtrace is from an older kernel.

I actually did update the backtrace in V3 with the one from next 20260413 that contained b44596ffe1b4 ("ARM: Allow to enable RT") from stable-rt/v6.12-rt-rebase branch [1] .

I think I misunderstood the usage of "softirq is raised" vs. "softirq is invoked" above . Is it possible that there was an already raised softirq before the threaded IRQ handler was invoked, and __netdev_alloc_skb() is what invoked that softirq ?

It is not impossible. Something needs to netif_wake_queue() and ks8851_irq() must only report IRQ_RXI (not IRQ_TXI). Then it can happen. But usually the driver "stops" the queue if it can't process any new packets and resumes it once a packet has been sent so it has room again.

Quoted 7 lines
Quoted 4 lines

If there is a flaw in my the theory please explain _how_ you managed that get that backtrace. I am sure it must have from an older kernel and _now_ this lockup also happens on !RT kernels (except for the SPI platform).

I used [1] , with PREEMPT_RT enabled , on stm32mp157c SoC . I ran iperf3 -s on the stm32 side, iperf3 -c 192.168.1.2 -t 0 --bidir on the hostpc side. The backtrace happened shortly after.

Hmm. Let me accept it then.

Sebastian

#10 lore raw
Quoted 13 lines · Sebastian Andrzej Siewior
Quoted 11 lines · Marek Vasut
Quoted 2 lines

memory allocation. Therefore I am saying this backtrace is from an older kernel.

I actually did update the backtrace in V3 with the one from next 20260413 that contained b44596ffe1b4 ("ARM: Allow to enable RT") from stable-rt/v6.12-rt-rebase branch [1] .

I think I misunderstood the usage of "softirq is raised" vs. "softirq is invoked" above . Is it possible that there was an already raised softirq before the threaded IRQ handler was invoked, and __netdev_alloc_skb() is what invoked that softirq ?

Quoted 4 lines

It is not impossible. Something needs to netif_wake_queue() and ks8851_irq() must only report IRQ_RXI (not IRQ_TXI). Then it can happen. But usually the driver "stops" the queue if it can't process any new packets and resumes it once a packet has been sent so it has room again.

This driver .start_xmit is very simple, if there is space in the 6 kiB TX FIFO, then the packet is written into it, otherwise the .start_xmit returns NETDEV_TX_BUSY . There does not seem to be any netif_{start,stop,wake}_queue() in the .start_xmit path.

maintainer
#12 lore raw
Quoted 7 lines · On Tue, 14 Apr 2026 08:09:31 -0700 Jakub Kicinski wrote:
Quoted 1 line · On Tue, 14 Apr 2026 14:57:53 +0200 Sebastian Andrzej Siewior wrote:

Maybe I'm not being forceful enough.

Putting workarounds in the drivers is unacceptable. __netdev_alloc_skb() must be legal to call under an _irq spin lock.

My bad, only read your reply to the old thread now.