net/802/tr: use interrupt-safe locking

PATCH Applied John W. Linville 4 participants 5 messages Aug 17, 2005 – Aug 21, 2005 1 message not in the archive

Subsystems (2)

Networking Network drivers

lore

Share & tools

lore

John W. Linville

author

2005-08-17 20:49 UTC #0 lore raw

PATCH 1/1 net/802/tr: use interrupt-safe locking

Change operations on rif_lock from spin_{un}lock_bh to spin_{un}lock_irq{save,restore} equivalents. Some of the rif_lock critical sections are called from interrupt context via tr_type_trans->tr_add_rif_info. The TR NIC drivers call tr_type_trans from their packet receive handlers.

Signed-off-by Jay Vosburg <[email protected]>

Signed-off-by John W. Linville <[email protected]>

---
It is my understanding that this same patch has been submitted multiple times in the past. Some of those submissions were around a year ago, but it does not seem to have been committed.

FWIW, this patch is currently being carried in the Fedora and RHEL kernels. It certainly looks like it is necessary to me. Can we get some movement on this?

 net/802/tr.c |   22 ++++++++++++----------
 1 files changed, 12 insertions(+), 10 deletions(-)

net/802/tr.c +12 ~~-10~~

@@ -251,10 +251,11 @@ void tr_source_route(struct sk_buff *skb
 	unsigned int hash;
 	struct rif_cache *entry;
 	unsigned char *olddata;
+	unsigned long flags;
 	static const unsigned char mcast_func_addr[] 
 		= {0xC0,0x00,0x00,0x04,0x00,0x00};
 	
-	spin_lock_bh(&rif_lock);
+	spin_lock_irqsave(&rif_lock, flags);
 
 	/*
 	 *	Broadcasts are single route as stated in RFC 1042

@@ -323,7 +324,7 @@ printk("source routing for %02X:%02X:%02
 	else 
 		slack = 18 - ((ntohs(trh->rcf) & TR_RCF_LEN_MASK)>>8);
 	olddata = skb->data;
-	spin_unlock_bh(&rif_lock);
+	spin_unlock_irqrestore(&rif_lock, flags);
 
 	skb_pull(skb, slack);
 	memmove(skb->data, olddata, sizeof(struct trh_hdr) - slack);

@@ -337,10 +338,11 @@ printk("source routing for %02X:%02X:%02
 static void tr_add_rif_info(struct trh_hdr *trh, struct net_device *dev)
 {
 	unsigned int hash, rii_p = 0;
+	unsigned long flags;
 	struct rif_cache *entry;
 
 
-	spin_lock_bh(&rif_lock);
+	spin_lock_irqsave(&rif_lock, flags);
 	
 	/*
 	 *	Firstly see if the entry exists

@@ -378,7 +380,7 @@ printk("adding rif_entry: addr:%02X:%02X
 		if(!entry) 
 		{
 			printk(KERN_DEBUG "tr.c: Couldn't malloc rif cache entry !\n");
-			spin_unlock_bh(&rif_lock);
+			spin_unlock_irqrestore(&rif_lock, flags);
 			return;
 		}

@@ -420,7 +422,7 @@ printk("updating rif_entry: addr:%02X:%0
 		    }                                         
            	entry->last_used=jiffies;               
 	}
-	spin_unlock_bh(&rif_lock);
+	spin_unlock_irqrestore(&rif_lock, flags);
 }
 
 /*

@@ -430,9 +432,9 @@ printk("updating rif_entry: addr:%02X:%0
 static void rif_check_expire(unsigned long dummy) 
 {
 	int i;
-	unsigned long next_interval = jiffies + sysctl_tr_rif_timeout/2;
+	unsigned long flags, next_interval = jiffies + sysctl_tr_rif_timeout/2;
 
-	spin_lock_bh(&rif_lock);
+	spin_lock_irqsave(&rif_lock, flags);
 	
 	for(i =0; i < RIF_TABLE_SIZE; i++) {
 		struct rif_cache *entry, **pentry;

@@ -454,7 +456,7 @@ static void rif_check_expire(unsigned lo
 		}
 	}
 	
-	spin_unlock_bh(&rif_lock);
+	spin_unlock_irqrestore(&rif_lock, flags);
 
 	mod_timer(&rif_timer, next_interval);

@@ -485,7 +487,7 @@ static struct rif_cache *rif_get_idx(lof
 
 static void *rif_seq_start(struct seq_file *seq, loff_t *pos)
 {
-	spin_lock_bh(&rif_lock);
+	spin_lock_irq(&rif_lock);
 
 	return *pos ? rif_get_idx(*pos - 1) : SEQ_START_TOKEN;
 }

@@ -516,7 +518,7 @@ static void *rif_seq_next(struct seq_fil
 
 static void rif_seq_stop(struct seq_file *seq, void *v)
 {
-	spin_unlock_bh(&rif_lock);
+	spin_unlock_irq(&rif_lock);
 }
 
 static int rif_seq_show(struct seq_file *seq, void *v)

David S. Miller

maintainer Applied

2005-08-17 22:00 UTC #1 lore raw

From: "John W. Linville" <[email protected]>
Date: Wed, 17 Aug 2005 16:49:59 -0400

Quoted 5 lines

Change operations on rif_lock from spin_{un}lock_bh to spin_{un}lock_irq{save,restore} equivalents. Some of the rif_lock critical sections are called from interrupt context via tr_type_trans->tr_add_rif_info. The TR NIC drivers call tr_type_trans from their packet receive handlers.

Applied, I'll try to get this into 2.6.13, but it may have to wait for 2.6.14

A message that is not in the archive (sent off-list or not yet received): [email protected]

Jay Vosburgh

2005-08-17 21:48 UTC #3 lore raw

Quoted 7 lines · John W. Linville

Change operations on rif_lock from spin_{un}lock_bh to spin_{un}lock_irq{save,restore} equivalents. Some of the rif_lock critical sections are called from interrupt context via tr_type_trans->tr_add_rif_info. The TR NIC drivers call tr_type_trans from their packet receive handlers.
Signed-off-by Jay Vosburg <[email protected]>

Pretty close.

Signed-off-by Jay Vosburgh <[email protected]>

Quoted 3 lines

It is my understanding that this same patch has been submitted multiple times in the past. Some of those submissions were around a year ago, but it does not seem to have been committed.

I believe that I originally wrote and posted this patch in the
appended message; I recall posting it a few times in various places.

Quoted 3 lines

FWIW, this patch is currently being carried in the Fedora and RHEL kernels. It certainly looks like it is necessary to me. Can we get some movement on this?

	It's in the SuSE kernel as well.

	-J

---
-Jay Vosburgh, IBM Linux Technology Center, [email protected]

To: Paul Mackerras <[email protected]>
Cc: [email protected], [email protected]
Subject: Re: spin_lock_bh() called in irq handler
Date: Wed, 28 Apr 2004 10:55:33 -0700
From: Jay Vosburgh <[email protected]>

Quoted 7 lines

I had a look and found that all of the token-ring drivers call tr_type_trans() at interrupt level. That seems perfectly reasonable to me. To fix the bug, it seems to me that there are two options:
either move the tr_add_rif_info() call elsewhere (but I have no idea where) or else use spin_lock_irqsave instead of spin_lock_bh.
Which is the more appropriate fix?

I'm guessing spin_lock_irqsave; would the following be
appropriate? I'm not absolutely sure about using spin_(un)lock_irq in rif_seq_start/stop, but it'd be complicated to deal with the flags in that case.

I've built this and given it some basic testing, but not really
hammered on it. The system doesn't panic when I cat /proc/net/tr_rif, which is a good sign.

-J

---
-Jay Vosburgh, IBM Linux Technology Center, [email protected]

net/802/tr.c +12 ~~-10~~

diff -urN linux-2.6.5-orig/net/802/tr.c linux-2.6.5/net/802/tr.c
@@ -250,10 +250,11 @@
 	unsigned int hash;
 	struct rif_cache_s *entry;
 	unsigned char *olddata;
+	unsigned long flags;
 	static const unsigned char mcast_func_addr[] 
 		= {0xC0,0x00,0x00,0x04,0x00,0x00};
 	
-	spin_lock_bh(&rif_lock);
+	spin_lock_irqsave(&rif_lock, flags);
 
 	/*
 	 *	Broadcasts are single route as stated in RFC 1042

@@ -322,7 +323,7 @@
 	else 
 		slack = 18 - ((ntohs(trh->rcf) & TR_RCF_LEN_MASK)>>8);
 	olddata = skb->data;
-	spin_unlock_bh(&rif_lock);
+	spin_unlock_irqrestore(&rif_lock, flags);
 
 	skb_pull(skb, slack);
 	memmove(skb->data, olddata, sizeof(struct trh_hdr) - slack);

@@ -336,10 +337,11 @@
 static void tr_add_rif_info(struct trh_hdr *trh, struct net_device *dev)
 {
 	unsigned int hash, rii_p = 0;
+	unsigned long flags;
 	struct rif_cache_s *entry;
 
 
-	spin_lock_bh(&rif_lock);
+	spin_lock_irqsave(&rif_lock, flags);
 	
 	/*
 	 *	Firstly see if the entry exists

@@ -377,7 +379,7 @@
 		if(!entry) 
 		{
 			printk(KERN_DEBUG "tr.c: Couldn't malloc rif cache entry !\n");
-			spin_unlock_bh(&rif_lock);
+			spin_unlock_irqrestore(&rif_lock, flags);
 			return;
 		}

@@ -419,7 +421,7 @@
 		    }                                         
            	entry->last_used=jiffies;               
 	}
-	spin_unlock_bh(&rif_lock);
+	spin_unlock_irqrestore(&rif_lock, flags);
 }
 
 /*

@@ -429,9 +431,9 @@
 static void rif_check_expire(unsigned long dummy) 
 {
 	int i;
-	unsigned long next_interval = jiffies + sysctl_tr_rif_timeout/2;
+	unsigned long flags, next_interval = jiffies + sysctl_tr_rif_timeout/2;
 
-	spin_lock_bh(&rif_lock);
+	spin_lock_irqsave(&rif_lock, flags);
 	
 	for(i =0; i < RIF_TABLE_SIZE; i++) {
 		struct rif_cache_s *entry, **pentry;

@@ -453,7 +455,7 @@
 		}
 	}
 	
-	spin_unlock_bh(&rif_lock);
+	spin_unlock_irqrestore(&rif_lock, flags);
 
 	mod_timer(&rif_timer, next_interval);

@@ -484,7 +486,7 @@
 
 static void *rif_seq_start(struct seq_file *seq, loff_t *pos)
 {
-	spin_lock_bh(&rif_lock);
+	spin_lock_irq(&rif_lock);
 
 	return *pos ? rif_get_idx(*pos - 1) : SEQ_START_TOKEN;
 }

@@ -515,7 +517,7 @@
 
 static void rif_seq_stop(struct seq_file *seq, void *v)
 {
-	spin_unlock_bh(&rif_lock);
+	spin_unlock_irq(&rif_lock);
 }
 
 static int rif_seq_show(struct seq_file *seq, void *v)

-J

---
-Jay Vosburgh, IBM Linux Technology Center, [email protected]

John W. Linville

author

2005-08-18 01:55 UTC #4 lore raw

Quoted 7 lines · Jay Vosburgh

Quoted 1 line · John W. Linville
Signed-off-by Jay Vosburg <[email protected]>
Pretty close.
Signed-off-by Jay Vosburgh <[email protected]>

Ooops...sorry! Tired, sloppy typing... :-(

Quoted 2 lines

I believe that I originally wrote and posted this patch in the
appended message; I recall posting it a few times in various places.

That is my understanding as well, which is why I put (an unfortunately incorrect) S-o-b line for you ahead of mine. Should I have done something else?

Thanks,

John

Signature · 2 lines

John W. Linville
[email protected]

Andrew Morton

2005-08-21 08:38 UTC #5 lore raw

In reply to Jay Vosburgh

Quoted 6 lines · Jay Vosburgh

Quoted 4 lines
FWIW, this patch is currently being carried in the Fedora and RHEL kernels. It certainly looks like it is necessary to me. Can we get some movement on this?
It's in the SuSE kernel as well.

For how long has this fix been in the vendor kernels?

Could someone please tell us why there are unmerged bugfixes in vendor kernels?

Are there any more?