Skip to contentSkip to conversation

Reduce netfilter memory use on MP systems

PATCH In review Andi Kleen 4 participants 6 messages Feb 4, 2005 – Feb 7, 2005

Subsystems (4)
lore
Share & tools
lore

Andi Kleen

author
#0 lore raw

PATCH 1/1 Reduce netfilter memory use on MP systems

On kernels compiled with a big NR_CPUS netfilter rules would eat a lot of memory because all counters would be duplicated for all NR_CPUs CPUs. With NR_CPUS=256 this would add up to many MBs of memory.

This patch only allocates enough memory for the possible CPUs, which is usually a much smaller number than NR_CPUS.

This allows loading of bigger rule sets on 64bit systems. There is still a limit because someone else broke vmalloc to have a 64MB limit on 64bit systems for single allocations, 129MB on 32bit. It allocates an array of pages with kmalloc and kmalloc has a 128K limit. To be fixed with a separate patch.

64bit systems were hurt worst because they tend to have big NR_CPUS and the counters need more memory there, and the vmalloc limit is lower. But it will raise the limits even on 32bit.

And in general it saves a lot of memory.

Tested only on a small dual CPU box.

net/ipv4/netfilter/ip_tables.c renamed from net/ipv4/netfilter/ip_tables.c-o +5 -5
diff -u linux/net/ipv4/netfilter/ip_tables.c-o linux/net/ipv4/netfilter/ip_tables.c
@@ -923,7 +923,7 @@
 	}
 
 	/* And one copy for every other CPU */
-	for (i = 1; i < NR_CPUS; i++) {
+	for (i = 1; i < num_possible_cpus(); i++) {
 		memcpy(newinfo->entries + SMP_ALIGN(newinfo->size)*i,
 		       newinfo->entries,
 		       SMP_ALIGN(newinfo->size));
@@ -945,7 +945,7 @@
 		struct ipt_entry *table_base;
 		unsigned int i;
 
-		for (i = 0; i < NR_CPUS; i++) {
+		for (i = 0; i < num_possible_cpus(); i++) {
 			table_base =
 				(void *)newinfo->entries
 				+ TABLE_OFFSET(newinfo, i);
@@ -992,7 +992,7 @@
 	unsigned int cpu;
 	unsigned int i;
 
-	for (cpu = 0; cpu < NR_CPUS; cpu++) {
+	for (cpu = 0; cpu < num_possible_cpus(); cpu++) {
 		i = 0;
 		IPT_ENTRY_ITERATE(t->entries + TABLE_OFFSET(t, cpu),
 				  t->size,
@@ -1130,7 +1130,7 @@
 		return -ENOMEM;
 
 	newinfo = vmalloc(sizeof(struct ipt_table_info)
-			  + SMP_ALIGN(tmp.size) * NR_CPUS);
+			  + SMP_ALIGN(tmp.size) * num_possible_cpus());
 	if (!newinfo)
 		return -ENOMEM;
 
@@ -1460,7 +1460,7 @@
 		= { 0, 0, 0, { 0 }, { 0 }, { } };
 
 	newinfo = vmalloc(sizeof(struct ipt_table_info)
-			  + SMP_ALIGN(repl->size) * NR_CPUS);
+			  + SMP_ALIGN(repl->size) * num_possible_cpus());
 	if (!newinfo)
 		return -ENOMEM;
 
#1 lore raw
Quoted 14 lines · Andi Kleen

On kernels compiled with a big NR_CPUS netfilter rules would eat a lot of memory because all counters would be duplicated for all NR_CPUs CPUs. With NR_CPUS=256 this would add up to many MBs of memory.

This patch only allocates enough memory for the possible CPUs, which is usually a much smaller number than NR_CPUS.

This allows loading of bigger rule sets on 64bit systems. There is still a limit because someone else broke vmalloc to have a 64MB limit on 64bit systems for single allocations, 129MB on 32bit. It allocates an array of pages with kmalloc and kmalloc has a 128K limit. To be fixed with a separate patch.

Quoted 5 lines

64bit systems were hurt worst because they tend to have big NR_CPUS and the counters need more memory there, and the vmalloc limit is lower. But it will raise the limits even on 32bit.

And in general it saves a lot of memory.

Patrick, could you apply and submit this patch to Davem? Or if Davem applies it himself. It's pretty obvious and would help small SMP machines with distribution kernels and/or strange admins.

Signature · 1 line
/Martin
author
#2 lore raw
Quoted 5 lines
Quoted 1 line

And in general it saves a lot of memory.

Patrick, could you apply and submit this patch to Davem? Or if Davem applies it himself. It's pretty obvious and would help small SMP machines with distribution kernels and/or strange admins.

The main motivation is actually not to save the memory (that's just a useful side effect), but increase the max limit on 64bit systems. Fixing it fully will require fixing vmalloc of course, but it already help. Without it you can't get more than ~3800 rules
on a 64bit system with NR_CPUS==128 and 128 byte cache lines.

-Andi

#3 lore raw
Quoted 6 lines · Andi Kleen

The main motivation is actually not to save the memory (that's just a useful side effect), but increase the max limit on 64bit systems. Fixing it fully will require fixing vmalloc of course, but it already help. Without it you can't get more than ~3800 rules
on a 64bit system with NR_CPUS==128 and 128 byte cache lines.

Thanks Andi, I've added the patch to my 2.6.12 tree. I've also made the same change in arp_tables, ip6_tables and ebtables for consistency.

Regards
Patrick

===== net/bridge/netfilter/ebtables.c 1.17 vs edited =====

net/bridge/netfilter/ebtables.c +9 -9
@@ -822,10 +822,10 @@
 		/* this will get free'd in do_replace()/ebt_register_table()
 		   if an error occurs */
 		newinfo->chainstack = (struct ebt_chainstack **)
-		   vmalloc(NR_CPUS * sizeof(struct ebt_chainstack));
+		   vmalloc(num_possible_cpus() * sizeof(struct ebt_chainstack));
 		if (!newinfo->chainstack)
 			return -ENOMEM;
-		for (i = 0; i < NR_CPUS; i++) {
+		for (i = 0; i < num_possible_cpus(); i++) {
 			newinfo->chainstack[i] =
 			   vmalloc(udc_cnt * sizeof(struct ebt_chainstack));
 			if (!newinfo->chainstack[i]) {
@@ -898,7 +898,7 @@
 	memcpy(counters, oldcounters,
 	   sizeof(struct ebt_counter) * nentries);
 	/* add other counters to those of cpu 0 */
-	for (cpu = 1; cpu < NR_CPUS; cpu++) {
+	for (cpu = 1; cpu < num_possible_cpus(); cpu++) {
 		counter_base = COUNTER_BASE(oldcounters, nentries, cpu);
 		for (i = 0; i < nentries; i++) {
 			counters[i].pcnt += counter_base[i].pcnt;
@@ -930,7 +930,7 @@
 		BUGPRINT("Entries_size never zero\n");
 		return -EINVAL;
 	}
-	countersize = COUNTER_OFFSET(tmp.nentries) * NR_CPUS;
+	countersize = COUNTER_OFFSET(tmp.nentries) * num_possible_cpus();
 	newinfo = (struct ebt_table_info *)
 	   vmalloc(sizeof(struct ebt_table_info) + countersize);
 	if (!newinfo)
@@ -1023,7 +1023,7 @@
 
 	vfree(table->entries);
 	if (table->chainstack) {
-		for (i = 0; i < NR_CPUS; i++)
+		for (i = 0; i < num_possible_cpus(); i++)
 			vfree(table->chainstack[i]);
 		vfree(table->chainstack);
 	}
@@ -1043,7 +1043,7 @@
 		vfree(counterstmp);
 	/* can be initialized in translate_table() */
 	if (newinfo->chainstack) {
-		for (i = 0; i < NR_CPUS; i++)
+		for (i = 0; i < num_possible_cpus(); i++)
 			vfree(newinfo->chainstack[i]);
 		vfree(newinfo->chainstack);
 	}
@@ -1137,7 +1137,7 @@
 		return -EINVAL;
 	}
 
-	countersize = COUNTER_OFFSET(table->table->nentries) * NR_CPUS;
+	countersize = COUNTER_OFFSET(table->table->nentries) * num_possible_cpus();
 	newinfo = (struct ebt_table_info *)
 	   vmalloc(sizeof(struct ebt_table_info) + countersize);
 	ret = -ENOMEM;
@@ -1191,7 +1191,7 @@
 	up(&ebt_mutex);
 free_chainstack:
 	if (newinfo->chainstack) {
-		for (i = 0; i < NR_CPUS; i++)
+		for (i = 0; i < num_possible_cpus(); i++)
 			vfree(newinfo->chainstack[i]);
 		vfree(newinfo->chainstack);
 	}
@@ -1215,7 +1215,7 @@
 	if (table->private->entries)
 		vfree(table->private->entries);
 	if (table->private->chainstack) {
-		for (i = 0; i < NR_CPUS; i++)
+		for (i = 0; i < num_possible_cpus(); i++)
 			vfree(table->private->chainstack[i]);
 		vfree(table->private->chainstack);
 	}

===== net/ipv4/netfilter/arp_tables.c 1.23 vs edited =====

net/ipv4/netfilter/arp_tables.c +4 -4
@@ -717,7 +717,7 @@
 	}
 
 	/* And one copy for every other CPU */
-	for (i = 1; i < NR_CPUS; i++) {
+	for (i = 1; i < num_possible_cpus(); i++) {
 		memcpy(newinfo->entries + SMP_ALIGN(newinfo->size)*i,
 		       newinfo->entries,
 		       SMP_ALIGN(newinfo->size));
@@ -768,7 +768,7 @@
 	unsigned int cpu;
 	unsigned int i;
 
-	for (cpu = 0; cpu < NR_CPUS; cpu++) {
+	for (cpu = 0; cpu < num_possible_cpus(); cpu++) {
 		i = 0;
 		ARPT_ENTRY_ITERATE(t->entries + TABLE_OFFSET(t, cpu),
 				   t->size,
@@ -886,7 +886,7 @@
 		return -ENOMEM;
 
 	newinfo = vmalloc(sizeof(struct arpt_table_info)
-			  + SMP_ALIGN(tmp.size) * NR_CPUS);
+			  + SMP_ALIGN(tmp.size) * num_possible_cpus());
 	if (!newinfo)
 		return -ENOMEM;
 
@@ -1159,7 +1159,7 @@
 		= { 0, 0, 0, { 0 }, { 0 }, { } };
 
 	newinfo = vmalloc(sizeof(struct arpt_table_info)
-			  + SMP_ALIGN(repl->size) * NR_CPUS);
+			  + SMP_ALIGN(repl->size) * num_possible_cpus());
 	if (!newinfo) {
 		ret = -ENOMEM;
 		return ret;

===== net/ipv6/netfilter/ip6_tables.c 1.39 vs edited =====

net/ipv6/netfilter/ip6_tables.c +5 -5
@@ -952,7 +952,7 @@
 	}
 
 	/* And one copy for every other CPU */
-	for (i = 1; i < NR_CPUS; i++) {
+	for (i = 1; i < num_possible_cpus(); i++) {
 		memcpy(newinfo->entries + SMP_ALIGN(newinfo->size)*i,
 		       newinfo->entries,
 		       SMP_ALIGN(newinfo->size));
@@ -974,7 +974,7 @@
 		struct ip6t_entry *table_base;
 		unsigned int i;
 
-		for (i = 0; i < NR_CPUS; i++) {
+		for (i = 0; i < num_possible_cpus(); i++) {
 			table_base =
 				(void *)newinfo->entries
 				+ TABLE_OFFSET(newinfo, i);
@@ -1021,7 +1021,7 @@
 	unsigned int cpu;
 	unsigned int i;
 
-	for (cpu = 0; cpu < NR_CPUS; cpu++) {
+	for (cpu = 0; cpu < num_possible_cpus(); cpu++) {
 		i = 0;
 		IP6T_ENTRY_ITERATE(t->entries + TABLE_OFFSET(t, cpu),
 				  t->size,
@@ -1155,7 +1155,7 @@
 		return -ENOMEM;
 
 	newinfo = vmalloc(sizeof(struct ip6t_table_info)
-			  + SMP_ALIGN(tmp.size) * NR_CPUS);
+			  + SMP_ALIGN(tmp.size) * num_possible_cpus());
 	if (!newinfo)
 		return -ENOMEM;
 
@@ -1469,7 +1469,7 @@
 		= { 0, 0, 0, { 0 }, { 0 }, { } };
 
 	newinfo = vmalloc(sizeof(struct ip6t_table_info)
-			  + SMP_ALIGN(repl->size) * NR_CPUS);
+			  + SMP_ALIGN(repl->size) * num_possible_cpus());
 	if (!newinfo)
 		return -ENOMEM;
 
#4 lore raw
In reply to Andi Kleen
Quoted 5 lines · Andi Kleen

On kernels compiled with a big NR_CPUS netfilter rules would eat a lot of memory because all counters would be duplicated for all NR_CPUs CPUs. With NR_CPUS=256 this would add up to many MBs of memory.

Thanks, Andi. I think the NR_CPUS is actually a remnescant of 2.3.x times when we didn't have num_possible_cpus() yet.

Also wrt. your vmalloc issues, I think there are floating around some patches which replace our vmalloc use by __get_free_pages() anyway.

Signature · 5 lines
- Harald Welte <[email protected]>             http://www.netfilter.org/
============================================================================
  "Fragmentation is like classful addressing -- an interesting early
   architectural error that shows how much experimentation was going
   on while IP was being designed."                    -- Paul Vixie
author
#5 lore raw
Quoted 2 lines

Also wrt. your vmalloc issues, I think there are floating around some patches which replace our vmalloc use by __get_free_pages() anyway.

Don't think they're needed (unless you want it for other reasons). The vmalloc limit is clearly a bug in itself and will be surely fixed.

-Andi