Re: net_tx_action race condition?

Re: net_tx_action race condition?

[Eric Dumazet]

On 03/28/2018 12:30 AM, Saurabh Kr wrote:
> Hi Eric/Angelo,
>  
> We are seeing the assertion error  in linux kernel 2.4.29  “*kernel: KERNEL: assertion (atomic_read(&skb->users) == 0) failed at dev.c(1397)**”.* Based on patch provided (_https://patchwork.kernel.org/patch/5368051/_ ) we merged the changes in linux kernel 2.4.29 but we are still facing the assertion error at dev.c (1397). Please let me know your thoughts.
>  
> *Before Merge**(linux 2.4.29)*
> ---------------------------------
>  
> static void net_tx_action(struct softirq_action *h)
> {
>         int cpu = smp_processor_id();
>  
>         if (softnet_data[cpu].completion_queue) {
>                 struct sk_buff *clist;
>  
>                 local_irq_disable();
>                 clist = softnet_data[cpu].completion_queue; // Existing code
>                 softnet_data[cpu].completion_queue = NULL;
>                 local_irq_enable();
>  
>                 while (clist != NULL) {
>                         struct sk_buff *skb = clist;
>                         clist = clist->next;
>  
>                         BUG_TRAP(atomic_read(&skb->users) == 0);
>                         __kfree_skb(skb);
>                 }
>         }
>  
>          ---------
>  
> *After Merge the changes based on available patch**(linux 2.4.29)**:*
> ------------------------------------------------------------------------------
>  
> static void net_tx_action(struct softirq_action *h)
> {
>         int cpu = smp_processor_id();
>  
>         if (softnet_data[cpu].completion_queue) {
>                 struct sk_buff *clist;
>  
>                 local_irq_disable();
>                 clist = *(volatile typeof(softnet_data[cpu].completion_queue) *)&( softnet_data[cpu].completion_queue);  // Modified line based on available patch
>                 softnet_data[cpu].completion_queue = NULL;
>                 local_irq_enable();
>  
>                 while (clist != NULL) {
>                         struct sk_buff *skb = clist;
>                         clist = clist->next;
>  
>                         BUG_TRAP(atomic_read(&skb->users) == 0);
>                         __kfree_skb(skb);
>                 }
>         }
>   ………….
>  
> Thanks & regards,
> Saurabh
>  

Thats simply prove (again) that this 'fix' was not the proper one.

I have no idea what is wrong, and there is no way I am going to look at 2.4.29 kernel...

net_tx_action race condition?

[Angelo Rizzi]

[-- Attachment #1: Type: text/plain, Size: 936 bytes --]

Hi Daniel,
Here attached the patch file you required.
The problem i've found is on the declaration of 'struct softnet_data 
*sd' in function 'net_tx_action'
What happens to me (i have an embedded system based on FPGA and a NIOS2 
cpu) is that, due to compiler optimization, the content of 
'sd->completion_queue' is saved in a CPU register before interrupt 
disabling (when the instruction 'if (sd->completion_queue) {' is 
executed) and then the register contents is used for interrupt-disabled 
assignment ('clist = sd->completion_queue') instead of re-read the 
variable contents.
This seems to lead to a race condition when an interrupt modifies the 
content of 'sd->completion_queue' between these two instructions.
What i have done to avoid this situation is to change the declaration of 
'struct softnet_data *sd' to 'volatile struct softnet_data *sd' and now 
everything seems to be ok.
I hope this will help.

Regards,
Angelo


[-- Attachment #2: net_tx_action_patch --]
[-- Type: text/plain, Size: 438 bytes --]

--- linux-2.6.32.64/net/core/dev.c.orig	2014-11-24 12:52:36.103335551 +0100
+++ linux-2.6.32.64/net/core/dev.c	2014-11-24 12:52:58.863949715 +0100
@@ -2069,7 +2069,7 @@ EXPORT_SYMBOL(netif_rx_ni);
 
 static void net_tx_action(struct softirq_action *h)
 {
-	struct softnet_data *sd = &__get_cpu_var(softnet_data);
+	volatile struct softnet_data *sd = &__get_cpu_var(softnet_data);
 
 	if (sd->completion_queue) {
 		struct sk_buff *clist;

Re: net_tx_action race condition?

[Herbert Xu]

Angelo Rizzi <angelo.rizzi@3dautomazione.it> wrote:
> [-- text/plain, encoding 7bit, charset: iso-8859-15, 22 lines --]
> 
> Hi Daniel,
> Here attached the patch file you required.
> The problem i've found is on the declaration of 'struct softnet_data 
> *sd' in function 'net_tx_action'
> What happens to me (i have an embedded system based on FPGA and a NIOS2 
> cpu) is that, due to compiler optimization, the content of 
> 'sd->completion_queue' is saved in a CPU register before interrupt 
> disabling (when the instruction 'if (sd->completion_queue) {' is 
> executed) and then the register contents is used for interrupt-disabled 
> assignment ('clist = sd->completion_queue') instead of re-read the 
> variable contents.
> This seems to lead to a race condition when an interrupt modifies the 
> content of 'sd->completion_queue' between these two instructions.
> What i have done to avoid this situation is to change the declaration of 
> 'struct softnet_data *sd' to 'volatile struct softnet_data *sd' and now 
> everything seems to be ok.
> I hope this will help.

local_irq_disable is supposed to contain a compiler barrier.  So
this is either a buggy arch or a buggy compiler.

Cheers,
-- 
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: net_tx_action race condition?

[Eric Dumazet]

On Mon, 2014-11-24 at 14:29 +0100, Angelo Rizzi wrote:
> Hi Daniel,
> Here attached the patch file you required.
> The problem i've found is on the declaration of 'struct softnet_data 
> *sd' in function 'net_tx_action'
> What happens to me (i have an embedded system based on FPGA and a NIOS2 
> cpu) is that, due to compiler optimization, the content of 
> 'sd->completion_queue' is saved in a CPU register before interrupt 
> disabling (when the instruction 'if (sd->completion_queue) {' is 
> executed) and then the register contents is used for interrupt-disabled 
> assignment ('clist = sd->completion_queue') instead of re-read the 
> variable contents.
> This seems to lead to a race condition when an interrupt modifies the 
> content of 'sd->completion_queue' between these two instructions.
> What i have done to avoid this situation is to change the declaration of 
> 'struct softnet_data *sd' to 'volatile struct softnet_data *sd' and now 
> everything seems to be ok.
> I hope this will help.
> 
> Regards,
> Angelo
> 

Do not add volatile in the kernel, this is not how we solve this kind of
problems.  Documentation/memory-barriers.txt and

Documentation/00-INDEX:476:volatile-considered-harmful.txt
Documentation/00-INDEX:477:     - Why the "volatile" type class should not be used


I am surprised this patch is needed. Many other 'bugs' would need
similar fixes.

local_irq_disable() MUST have a memory barrier. This looks like a bug in
one particular arch implementation.


On x86 for example, native_irq_disable() is really :

	asm volatile("cli": : :"memory");


diff --git a/net/core/dev.c b/net/core/dev.c
index ac4836241a965a71952469ba054f87d8dfca2d32..fa73072e515aa07fa8ae1bc39174b7d59c7a31a5 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -3406,7 +3406,7 @@ static void net_tx_action(struct softirq_action *h)
 		struct sk_buff *clist;
 
 		local_irq_disable();
-		clist = sd->completion_queue;
+		clist = ACCESS_ONCE(sd->completion_queue);
 		sd->completion_queue = NULL;
 		local_irq_enable();
 

Re: net_tx_action race condition?

[Angelo Rizzi]

Hi Eric,
Thanks for your reply
You are right, it seems a bug in the NIOS2 architecture port.
I will check how local_irq_disable()/local_irq_enable() is implemented 
on this kind of architecture.

Regards,
Angelo

Il 24/11/2014 16:33, Eric Dumazet ha scritto:
> On Mon, 2014-11-24 at 14:29 +0100, Angelo Rizzi wrote:
>> Hi Daniel,
>> Here attached the patch file you required.
>> The problem i've found is on the declaration of 'struct softnet_data
>> *sd' in function 'net_tx_action'
>> What happens to me (i have an embedded system based on FPGA and a NIOS2
>> cpu) is that, due to compiler optimization, the content of
>> 'sd->completion_queue' is saved in a CPU register before interrupt
>> disabling (when the instruction 'if (sd->completion_queue) {' is
>> executed) and then the register contents is used for interrupt-disabled
>> assignment ('clist = sd->completion_queue') instead of re-read the
>> variable contents.
>> This seems to lead to a race condition when an interrupt modifies the
>> content of 'sd->completion_queue' between these two instructions.
>> What i have done to avoid this situation is to change the declaration of
>> 'struct softnet_data *sd' to 'volatile struct softnet_data *sd' and now
>> everything seems to be ok.
>> I hope this will help.
>>
>> Regards,
>> Angelo
>>
> Do not add volatile in the kernel, this is not how we solve this kind of
> problems.  Documentation/memory-barriers.txt and
>
> Documentation/00-INDEX:476:volatile-considered-harmful.txt
> Documentation/00-INDEX:477:     - Why the "volatile" type class should not be used
>
>
> I am surprised this patch is needed. Many other 'bugs' would need
> similar fixes.
>
> local_irq_disable() MUST have a memory barrier. This looks like a bug in
> one particular arch implementation.
>
>
> On x86 for example, native_irq_disable() is really :
>
> 	asm volatile("cli": : :"memory");
>
>
> diff --git a/net/core/dev.c b/net/core/dev.c
> index ac4836241a965a71952469ba054f87d8dfca2d32..fa73072e515aa07fa8ae1bc39174b7d59c7a31a5 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -3406,7 +3406,7 @@ static void net_tx_action(struct softirq_action *h)
>   		struct sk_buff *clist;
>   
>   		local_irq_disable();
> -		clist = sd->completion_queue;
> +		clist = ACCESS_ONCE(sd->completion_queue);
>   		sd->completion_queue = NULL;
>   		local_irq_enable();
>   
>
>
>