[net] net: Make napi_hash_lock irq safe

Joe Damato posted 1 patch 1 year, 2 months ago
net/core/dev.c | 18 ++++++++++++------
1 file changed, 12 insertions(+), 6 deletions(-)
[net] net: Make napi_hash_lock irq safe
Posted by Joe Damato 1 year, 2 months ago
Make napi_hash_lock IRQ safe. It is used during the control path, and is
taken and released in napi_hash_add and napi_hash_del, which will
typically be called by calls to napi_enable and napi_disable.

This change avoids a deadlock in pcnet32 (and other any other drivers
which follow the same pattern):

 CPU 0:
 pcnet32_open
    spin_lock_irqsave(&lp->lock, ...)
      napi_enable
        napi_hash_add <- before this executes, CPU 1 proceeds
          spin_lock(napi_hash_lock)
       [...]
    spin_unlock_irqrestore(&lp->lock, flags);

 CPU 1:
   pcnet32_close
     napi_disable
       napi_hash_del
         spin_lock(napi_hash_lock)
          < INTERRUPT >
            pcnet32_interrupt
              spin_lock(lp->lock) <- DEADLOCK

Changing the napi_hash_lock to be IRQ safe prevents the IRQ from firing
on CPU 1 until napi_hash_lock is released, preventing the deadlock.

Cc: stable@vger.kernel.org
Fixes: 86e25f40aa1e ("net: napi: Add napi_config")
Reported-by: Guenter Roeck <linux@roeck-us.net>
Closes: https://lore.kernel.org/netdev/85dd4590-ea6b-427d-876a-1d8559c7ad82@roeck-us.net/
Suggested-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Joe Damato <jdamato@fastly.com>
---
 net/core/dev.c | 18 ++++++++++++------
 1 file changed, 12 insertions(+), 6 deletions(-)

diff --git a/net/core/dev.c b/net/core/dev.c
index 13d00fc10f55..45a8c3dd4a64 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -6557,18 +6557,22 @@ static void __napi_hash_add_with_id(struct napi_struct *napi,
 static void napi_hash_add_with_id(struct napi_struct *napi,
 				  unsigned int napi_id)
 {
-	spin_lock(&napi_hash_lock);
+	unsigned long flags;
+
+	spin_lock_irqsave(&napi_hash_lock, flags);
 	WARN_ON_ONCE(napi_by_id(napi_id));
 	__napi_hash_add_with_id(napi, napi_id);
-	spin_unlock(&napi_hash_lock);
+	spin_unlock_irqrestore(&napi_hash_lock, flags);
 }
 
 static void napi_hash_add(struct napi_struct *napi)
 {
+	unsigned long flags;
+
 	if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state))
 		return;
 
-	spin_lock(&napi_hash_lock);
+	spin_lock_irqsave(&napi_hash_lock, flags);
 
 	/* 0..NR_CPUS range is reserved for sender_cpu use */
 	do {
@@ -6578,7 +6582,7 @@ static void napi_hash_add(struct napi_struct *napi)
 
 	__napi_hash_add_with_id(napi, napi_gen_id);
 
-	spin_unlock(&napi_hash_lock);
+	spin_unlock_irqrestore(&napi_hash_lock, flags);
 }
 
 /* Warning : caller is responsible to make sure rcu grace period
@@ -6586,11 +6590,13 @@ static void napi_hash_add(struct napi_struct *napi)
  */
 static void napi_hash_del(struct napi_struct *napi)
 {
-	spin_lock(&napi_hash_lock);
+	unsigned long flags;
+
+	spin_lock_irqsave(&napi_hash_lock, flags);
 
 	hlist_del_init_rcu(&napi->napi_hash_node);
 
-	spin_unlock(&napi_hash_lock);
+	spin_unlock_irqrestore(&napi_hash_lock, flags);
 }
 
 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer)
-- 
2.25.1
Re: [net] net: Make napi_hash_lock irq safe
Posted by Eric Dumazet 1 year, 2 months ago
On Mon, Dec 2, 2024 at 7:21 PM Joe Damato <jdamato@fastly.com> wrote:
>
> Make napi_hash_lock IRQ safe. It is used during the control path, and is
> taken and released in napi_hash_add and napi_hash_del, which will
> typically be called by calls to napi_enable and napi_disable.
>
> This change avoids a deadlock in pcnet32 (and other any other drivers
> which follow the same pattern):
>
>  CPU 0:
>  pcnet32_open
>     spin_lock_irqsave(&lp->lock, ...)
>       napi_enable
>         napi_hash_add <- before this executes, CPU 1 proceeds
>           spin_lock(napi_hash_lock)
>        [...]
>     spin_unlock_irqrestore(&lp->lock, flags);
>
>  CPU 1:
>    pcnet32_close
>      napi_disable
>        napi_hash_del
>          spin_lock(napi_hash_lock)
>           < INTERRUPT >
>             pcnet32_interrupt
>               spin_lock(lp->lock) <- DEADLOCK
>
> Changing the napi_hash_lock to be IRQ safe prevents the IRQ from firing
> on CPU 1 until napi_hash_lock is released, preventing the deadlock.
>
> Cc: stable@vger.kernel.org
> Fixes: 86e25f40aa1e ("net: napi: Add napi_config")
> Reported-by: Guenter Roeck <linux@roeck-us.net>
> Closes: https://lore.kernel.org/netdev/85dd4590-ea6b-427d-876a-1d8559c7ad82@roeck-us.net/
> Suggested-by: Jakub Kicinski <kuba@kernel.org>
> Signed-off-by: Joe Damato <jdamato@fastly.com>

Reviewed-by: Eric Dumazet <edumazet@google.com>
Re: [net] net: Make napi_hash_lock irq safe
Posted by Guenter Roeck 1 year, 2 months ago
On Mon, Dec 02, 2024 at 06:21:02PM +0000, Joe Damato wrote:
> Make napi_hash_lock IRQ safe. It is used during the control path, and is
> taken and released in napi_hash_add and napi_hash_del, which will
> typically be called by calls to napi_enable and napi_disable.
> 
> This change avoids a deadlock in pcnet32 (and other any other drivers
> which follow the same pattern):
> 
>  CPU 0:
>  pcnet32_open
>     spin_lock_irqsave(&lp->lock, ...)
>       napi_enable
>         napi_hash_add <- before this executes, CPU 1 proceeds
>           spin_lock(napi_hash_lock)
>        [...]
>     spin_unlock_irqrestore(&lp->lock, flags);
> 
>  CPU 1:
>    pcnet32_close
>      napi_disable
>        napi_hash_del
>          spin_lock(napi_hash_lock)
>           < INTERRUPT >
>             pcnet32_interrupt
>               spin_lock(lp->lock) <- DEADLOCK
> 
> Changing the napi_hash_lock to be IRQ safe prevents the IRQ from firing
> on CPU 1 until napi_hash_lock is released, preventing the deadlock.
> 
> Cc: stable@vger.kernel.org
> Fixes: 86e25f40aa1e ("net: napi: Add napi_config")
> Reported-by: Guenter Roeck <linux@roeck-us.net>
> Closes: https://lore.kernel.org/netdev/85dd4590-ea6b-427d-876a-1d8559c7ad82@roeck-us.net/
> Suggested-by: Jakub Kicinski <kuba@kernel.org>
> Signed-off-by: Joe Damato <jdamato@fastly.com>

Tested-by: Guenter Roeck <linux@roeck-us.net>

> ---
>  net/core/dev.c | 18 ++++++++++++------
>  1 file changed, 12 insertions(+), 6 deletions(-)
> 
> diff --git a/net/core/dev.c b/net/core/dev.c
> index 13d00fc10f55..45a8c3dd4a64 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -6557,18 +6557,22 @@ static void __napi_hash_add_with_id(struct napi_struct *napi,
>  static void napi_hash_add_with_id(struct napi_struct *napi,
>  				  unsigned int napi_id)
>  {
> -	spin_lock(&napi_hash_lock);
> +	unsigned long flags;
> +
> +	spin_lock_irqsave(&napi_hash_lock, flags);
>  	WARN_ON_ONCE(napi_by_id(napi_id));
>  	__napi_hash_add_with_id(napi, napi_id);
> -	spin_unlock(&napi_hash_lock);
> +	spin_unlock_irqrestore(&napi_hash_lock, flags);
>  }
>  
>  static void napi_hash_add(struct napi_struct *napi)
>  {
> +	unsigned long flags;
> +
>  	if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state))
>  		return;
>  
> -	spin_lock(&napi_hash_lock);
> +	spin_lock_irqsave(&napi_hash_lock, flags);
>  
>  	/* 0..NR_CPUS range is reserved for sender_cpu use */
>  	do {
> @@ -6578,7 +6582,7 @@ static void napi_hash_add(struct napi_struct *napi)
>  
>  	__napi_hash_add_with_id(napi, napi_gen_id);
>  
> -	spin_unlock(&napi_hash_lock);
> +	spin_unlock_irqrestore(&napi_hash_lock, flags);
>  }
>  
>  /* Warning : caller is responsible to make sure rcu grace period
> @@ -6586,11 +6590,13 @@ static void napi_hash_add(struct napi_struct *napi)
>   */
>  static void napi_hash_del(struct napi_struct *napi)
>  {
> -	spin_lock(&napi_hash_lock);
> +	unsigned long flags;
> +
> +	spin_lock_irqsave(&napi_hash_lock, flags);
>  
>  	hlist_del_init_rcu(&napi->napi_hash_node);
>  
> -	spin_unlock(&napi_hash_lock);
> +	spin_unlock_irqrestore(&napi_hash_lock, flags);
>  }
>  
>  static enum hrtimer_restart napi_watchdog(struct hrtimer *timer)
> -- 
> 2.25.1
>