linux/include/net/dst.h

489 lines
12 KiB
C
Raw Normal View History

/*
* net/dst.h Protocol independent destination cache definitions.
*
* Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
*
*/
#ifndef _NET_DST_H
#define _NET_DST_H
#include <net/dst_ops.h>
#include <linux/netdevice.h>
#include <linux/rtnetlink.h>
#include <linux/rcupdate.h>
#include <linux/bug.h>
#include <linux/jiffies.h>
#include <net/neighbour.h>
#include <asm/processor.h>
#define DST_GC_MIN (HZ/10)
#define DST_GC_INC (HZ/2)
#define DST_GC_MAX (120*HZ)
/* Each dst_entry has reference count and sits in some parent list(s).
* When it is removed from parent list, it is "freed" (dst_free).
* After this it enters dead state (dst->obsolete > 0) and if its refcnt
* is zero, it can be destroyed immediately, otherwise it is added
* to gc list and garbage collector periodically checks the refcnt.
*/
struct sk_buff;
struct dst_entry {
struct rcu_head rcu_head;
struct dst_entry *child;
struct net_device *dev;
net: Implement read-only protection and COW'ing of metrics. Routing metrics are now copy-on-write. Initially a route entry points it's metrics at a read-only location. If a routing table entry exists, it will point there. Else it will point at the all zero metric place-holder called 'dst_default_metrics'. The writeability state of the metrics is stored in the low bits of the metrics pointer, we have two bits left to spare if we want to store more states. For the initial implementation, COW is implemented simply via kmalloc. However future enhancements will change this to place the writable metrics somewhere else, in order to increase sharing. Very likely this "somewhere else" will be the inetpeer cache. Note also that this means that metrics updates may transiently fail if we cannot COW the metrics successfully. But even by itself, this patch should decrease memory usage and increase cache locality especially for routing workloads. In those cases the read-only metric copies stay in place and never get written to. TCP workloads where metrics get updated, and those rare cases where PMTU triggers occur, will take a very slight performance hit. But that hit will be alleviated when the long-term writable metrics move to a more sharable location. Since the metrics storage went from a u32 array of RTAX_MAX entries to what is essentially a pointer, some retooling of the dst_entry layout was necessary. Most importantly, we need to preserve the alignment of the reference count so that it doesn't share cache lines with the read-mostly state, as per Eric Dumazet's alignment assertion checks. The only non-trivial bit here is the move of the 'flags' member into the writeable cacheline. This is OK since we are always accessing the flags around the same moment when we made a modification to the reference count. Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-27 04:51:05 +00:00
struct dst_ops *ops;
unsigned long _metrics;
ipv6: fix race condition regarding dst->expires and dst->from. Eric Dumazet wrote: | Some strange crashes happen in rt6_check_expired(), with access | to random addresses. | | At first glance, it looks like the RTF_EXPIRES and | stuff added in commit 1716a96101c49186b | (ipv6: fix problem with expired dst cache) | are racy : same dst could be manipulated at the same time | on different cpus. | | At some point, our stack believes rt->dst.from contains a dst pointer, | while its really a jiffie value (as rt->dst.expires shares the same area | of memory) | | rt6_update_expires() should be fixed, or am I missing something ? | | CC Neil because of https://bugzilla.redhat.com/show_bug.cgi?id=892060 Because we do not have any locks for dst_entry, we cannot change essential structure in the entry; e.g., we cannot change reference to other entity. To fix this issue, split 'from' and 'expires' field in dst_entry out of union. Once it is 'from' is assigned in the constructor, keep the reference until the very last stage of the life time of the object. Of course, it is unsafe to change 'from', so make rt6_set_from simple just for fresh entries. Reported-by: Eric Dumazet <eric.dumazet@gmail.com> Reported-by: Neil Horman <nhorman@tuxdriver.com> CC: Gao Feng <gaofeng@cn.fujitsu.com> Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org> Reviewed-by: Eric Dumazet <edumazet@google.com> Reported-by: Steinar H. Gunderson <sesse@google.com> Reviewed-by: Neil Horman <nhorman@tuxdriver.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-02-20 00:29:08 +00:00
unsigned long expires;
[NET]: Fix tbench regression in 2.6.25-rc1 Comparing with kernel 2.6.24, tbench result has regression with 2.6.25-rc1. 1) On 2 quad-core processor stoakley: 4%. 2) On 4 quad-core processor tigerton: more than 30%. bisect located below patch. b4ce92775c2e7ff9cf79cca4e0a19c8c5fd6287b is first bad commit commit b4ce92775c2e7ff9cf79cca4e0a19c8c5fd6287b Author: Herbert Xu <herbert@gondor.apana.org.au> Date: Tue Nov 13 21:33:32 2007 -0800 [IPV6]: Move nfheader_len into rt6_info The dst member nfheader_len is only used by IPv6. It's also currently creating a rather ugly alignment hole in struct dst. Therefore this patch moves it from there into struct rt6_info. Above patch changes the cache line alignment, especially member __refcnt. I did a testing by adding 2 unsigned long pading before lastuse, so the 3 members, lastuse/__refcnt/__use, are moved to next cache line. The performance is recovered. I created a patch to rearrange the members in struct dst_entry. With Eric and Valdis Kletnieks's suggestion, I made finer arrangement. 1) Move tclassid under ops in case CONFIG_NET_CLS_ROUTE=y. So sizeof(dst_entry)=200 no matter if CONFIG_NET_CLS_ROUTE=y/n. I tested many patches on my 16-core tigerton by moving tclassid to different place. It looks like tclassid could also have impact on performance. If moving tclassid before metrics, or just don't move tclassid, the performance isn't good. So I move it behind metrics. 2) Add comments before __refcnt. On 16-core tigerton: If CONFIG_NET_CLS_ROUTE=y, the result with below patch is about 18% better than the one without the patch; If CONFIG_NET_CLS_ROUTE=n, the result with below patch is about 30% better than the one without the patch. With 32bit 2.6.25-rc1 on 8-core stoakley, the new patch doesn't introduce regression. Thank Eric, Valdis, and David! Signed-off-by: Zhang Yanmin <yanmin.zhang@intel.com> Acked-by: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-03-13 05:52:37 +00:00
struct dst_entry *path;
ipv6: fix race condition regarding dst->expires and dst->from. Eric Dumazet wrote: | Some strange crashes happen in rt6_check_expired(), with access | to random addresses. | | At first glance, it looks like the RTF_EXPIRES and | stuff added in commit 1716a96101c49186b | (ipv6: fix problem with expired dst cache) | are racy : same dst could be manipulated at the same time | on different cpus. | | At some point, our stack believes rt->dst.from contains a dst pointer, | while its really a jiffie value (as rt->dst.expires shares the same area | of memory) | | rt6_update_expires() should be fixed, or am I missing something ? | | CC Neil because of https://bugzilla.redhat.com/show_bug.cgi?id=892060 Because we do not have any locks for dst_entry, we cannot change essential structure in the entry; e.g., we cannot change reference to other entity. To fix this issue, split 'from' and 'expires' field in dst_entry out of union. Once it is 'from' is assigned in the constructor, keep the reference until the very last stage of the life time of the object. Of course, it is unsafe to change 'from', so make rt6_set_from simple just for fresh entries. Reported-by: Eric Dumazet <eric.dumazet@gmail.com> Reported-by: Neil Horman <nhorman@tuxdriver.com> CC: Gao Feng <gaofeng@cn.fujitsu.com> Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org> Reviewed-by: Eric Dumazet <edumazet@google.com> Reported-by: Steinar H. Gunderson <sesse@google.com> Reviewed-by: Neil Horman <nhorman@tuxdriver.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-02-20 00:29:08 +00:00
struct dst_entry *from;
#ifdef CONFIG_XFRM
struct xfrm_state *xfrm;
#else
void *__pad1;
#endif
int (*input)(struct sk_buff *);
int (*output)(struct sk_buff *);
unsigned short flags;
#define DST_HOST 0x0001
#define DST_NOXFRM 0x0002
#define DST_NOPOLICY 0x0004
#define DST_NOHASH 0x0008
#define DST_NOCACHE 0x0010
#define DST_NOCOUNT 0x0020
net: introduce DST_NOPEER dst flag Chris Boot reported crashes occurring in ipv6_select_ident(). [ 461.457562] RIP: 0010:[<ffffffff812dde61>] [<ffffffff812dde61>] ipv6_select_ident+0x31/0xa7 [ 461.578229] Call Trace: [ 461.580742] <IRQ> [ 461.582870] [<ffffffff812efa7f>] ? udp6_ufo_fragment+0x124/0x1a2 [ 461.589054] [<ffffffff812dbfe0>] ? ipv6_gso_segment+0xc0/0x155 [ 461.595140] [<ffffffff812700c6>] ? skb_gso_segment+0x208/0x28b [ 461.601198] [<ffffffffa03f236b>] ? ipv6_confirm+0x146/0x15e [nf_conntrack_ipv6] [ 461.608786] [<ffffffff81291c4d>] ? nf_iterate+0x41/0x77 [ 461.614227] [<ffffffff81271d64>] ? dev_hard_start_xmit+0x357/0x543 [ 461.620659] [<ffffffff81291cf6>] ? nf_hook_slow+0x73/0x111 [ 461.626440] [<ffffffffa0379745>] ? br_parse_ip_options+0x19a/0x19a [bridge] [ 461.633581] [<ffffffff812722ff>] ? dev_queue_xmit+0x3af/0x459 [ 461.639577] [<ffffffffa03747d2>] ? br_dev_queue_push_xmit+0x72/0x76 [bridge] [ 461.646887] [<ffffffffa03791e3>] ? br_nf_post_routing+0x17d/0x18f [bridge] [ 461.653997] [<ffffffff81291c4d>] ? nf_iterate+0x41/0x77 [ 461.659473] [<ffffffffa0374760>] ? br_flood+0xfa/0xfa [bridge] [ 461.665485] [<ffffffff81291cf6>] ? nf_hook_slow+0x73/0x111 [ 461.671234] [<ffffffffa0374760>] ? br_flood+0xfa/0xfa [bridge] [ 461.677299] [<ffffffffa0379215>] ? nf_bridge_update_protocol+0x20/0x20 [bridge] [ 461.684891] [<ffffffffa03bb0e5>] ? nf_ct_zone+0xa/0x17 [nf_conntrack] [ 461.691520] [<ffffffffa0374760>] ? br_flood+0xfa/0xfa [bridge] [ 461.697572] [<ffffffffa0374812>] ? NF_HOOK.constprop.8+0x3c/0x56 [bridge] [ 461.704616] [<ffffffffa0379031>] ? nf_bridge_push_encap_header+0x1c/0x26 [bridge] [ 461.712329] [<ffffffffa037929f>] ? br_nf_forward_finish+0x8a/0x95 [bridge] [ 461.719490] [<ffffffffa037900a>] ? nf_bridge_pull_encap_header+0x1c/0x27 [bridge] [ 461.727223] [<ffffffffa0379974>] ? br_nf_forward_ip+0x1c0/0x1d4 [bridge] [ 461.734292] [<ffffffff81291c4d>] ? nf_iterate+0x41/0x77 [ 461.739758] [<ffffffffa03748cc>] ? __br_deliver+0xa0/0xa0 [bridge] [ 461.746203] [<ffffffff81291cf6>] ? nf_hook_slow+0x73/0x111 [ 461.751950] [<ffffffffa03748cc>] ? __br_deliver+0xa0/0xa0 [bridge] [ 461.758378] [<ffffffffa037533a>] ? NF_HOOK.constprop.4+0x56/0x56 [bridge] This is caused by bridge netfilter special dst_entry (fake_rtable), a special shared entry, where attaching an inetpeer makes no sense. Problem is present since commit 87c48fa3b46 (ipv6: make fragment identifications less predictable) Introduce DST_NOPEER dst flag and make sure ipv6_select_ident() and __ip_select_ident() fallback to the 'no peer attached' handling. Reported-by: Chris Boot <bootc@bootc.net> Tested-by: Chris Boot <bootc@bootc.net> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-12-22 04:15:53 +00:00
#define DST_NOPEER 0x0040
#define DST_FAKE_RTABLE 0x0080
#define DST_XFRM_TUNNEL 0x0100
#define DST_XFRM_QUEUE 0x0200
unsigned short pending_confirm;
net: Implement read-only protection and COW'ing of metrics. Routing metrics are now copy-on-write. Initially a route entry points it's metrics at a read-only location. If a routing table entry exists, it will point there. Else it will point at the all zero metric place-holder called 'dst_default_metrics'. The writeability state of the metrics is stored in the low bits of the metrics pointer, we have two bits left to spare if we want to store more states. For the initial implementation, COW is implemented simply via kmalloc. However future enhancements will change this to place the writable metrics somewhere else, in order to increase sharing. Very likely this "somewhere else" will be the inetpeer cache. Note also that this means that metrics updates may transiently fail if we cannot COW the metrics successfully. But even by itself, this patch should decrease memory usage and increase cache locality especially for routing workloads. In those cases the read-only metric copies stay in place and never get written to. TCP workloads where metrics get updated, and those rare cases where PMTU triggers occur, will take a very slight performance hit. But that hit will be alleviated when the long-term writable metrics move to a more sharable location. Since the metrics storage went from a u32 array of RTAX_MAX entries to what is essentially a pointer, some retooling of the dst_entry layout was necessary. Most importantly, we need to preserve the alignment of the reference count so that it doesn't share cache lines with the read-mostly state, as per Eric Dumazet's alignment assertion checks. The only non-trivial bit here is the move of the 'flags' member into the writeable cacheline. This is OK since we are always accessing the flags around the same moment when we made a modification to the reference count. Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-27 04:51:05 +00:00
short error;
/* A non-zero value of dst->obsolete forces by-hand validation
* of the route entry. Positive values are set by the generic
* dst layer to indicate that the entry has been forcefully
* destroyed.
*
* Negative values are used by the implementation layer code to
* force invocation of the dst_ops->check() method.
*/
net: Implement read-only protection and COW'ing of metrics. Routing metrics are now copy-on-write. Initially a route entry points it's metrics at a read-only location. If a routing table entry exists, it will point there. Else it will point at the all zero metric place-holder called 'dst_default_metrics'. The writeability state of the metrics is stored in the low bits of the metrics pointer, we have two bits left to spare if we want to store more states. For the initial implementation, COW is implemented simply via kmalloc. However future enhancements will change this to place the writable metrics somewhere else, in order to increase sharing. Very likely this "somewhere else" will be the inetpeer cache. Note also that this means that metrics updates may transiently fail if we cannot COW the metrics successfully. But even by itself, this patch should decrease memory usage and increase cache locality especially for routing workloads. In those cases the read-only metric copies stay in place and never get written to. TCP workloads where metrics get updated, and those rare cases where PMTU triggers occur, will take a very slight performance hit. But that hit will be alleviated when the long-term writable metrics move to a more sharable location. Since the metrics storage went from a u32 array of RTAX_MAX entries to what is essentially a pointer, some retooling of the dst_entry layout was necessary. Most importantly, we need to preserve the alignment of the reference count so that it doesn't share cache lines with the read-mostly state, as per Eric Dumazet's alignment assertion checks. The only non-trivial bit here is the move of the 'flags' member into the writeable cacheline. This is OK since we are always accessing the flags around the same moment when we made a modification to the reference count. Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-27 04:51:05 +00:00
short obsolete;
#define DST_OBSOLETE_NONE 0
#define DST_OBSOLETE_DEAD 2
#define DST_OBSOLETE_FORCE_CHK -1
#define DST_OBSOLETE_KILL -2
net: Implement read-only protection and COW'ing of metrics. Routing metrics are now copy-on-write. Initially a route entry points it's metrics at a read-only location. If a routing table entry exists, it will point there. Else it will point at the all zero metric place-holder called 'dst_default_metrics'. The writeability state of the metrics is stored in the low bits of the metrics pointer, we have two bits left to spare if we want to store more states. For the initial implementation, COW is implemented simply via kmalloc. However future enhancements will change this to place the writable metrics somewhere else, in order to increase sharing. Very likely this "somewhere else" will be the inetpeer cache. Note also that this means that metrics updates may transiently fail if we cannot COW the metrics successfully. But even by itself, this patch should decrease memory usage and increase cache locality especially for routing workloads. In those cases the read-only metric copies stay in place and never get written to. TCP workloads where metrics get updated, and those rare cases where PMTU triggers occur, will take a very slight performance hit. But that hit will be alleviated when the long-term writable metrics move to a more sharable location. Since the metrics storage went from a u32 array of RTAX_MAX entries to what is essentially a pointer, some retooling of the dst_entry layout was necessary. Most importantly, we need to preserve the alignment of the reference count so that it doesn't share cache lines with the read-mostly state, as per Eric Dumazet's alignment assertion checks. The only non-trivial bit here is the move of the 'flags' member into the writeable cacheline. This is OK since we are always accessing the flags around the same moment when we made a modification to the reference count. Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-27 04:51:05 +00:00
unsigned short header_len; /* more space at head required */
unsigned short trailer_len; /* space to reserve at tail */
#ifdef CONFIG_IP_ROUTE_CLASSID
[NET]: Fix tbench regression in 2.6.25-rc1 Comparing with kernel 2.6.24, tbench result has regression with 2.6.25-rc1. 1) On 2 quad-core processor stoakley: 4%. 2) On 4 quad-core processor tigerton: more than 30%. bisect located below patch. b4ce92775c2e7ff9cf79cca4e0a19c8c5fd6287b is first bad commit commit b4ce92775c2e7ff9cf79cca4e0a19c8c5fd6287b Author: Herbert Xu <herbert@gondor.apana.org.au> Date: Tue Nov 13 21:33:32 2007 -0800 [IPV6]: Move nfheader_len into rt6_info The dst member nfheader_len is only used by IPv6. It's also currently creating a rather ugly alignment hole in struct dst. Therefore this patch moves it from there into struct rt6_info. Above patch changes the cache line alignment, especially member __refcnt. I did a testing by adding 2 unsigned long pading before lastuse, so the 3 members, lastuse/__refcnt/__use, are moved to next cache line. The performance is recovered. I created a patch to rearrange the members in struct dst_entry. With Eric and Valdis Kletnieks's suggestion, I made finer arrangement. 1) Move tclassid under ops in case CONFIG_NET_CLS_ROUTE=y. So sizeof(dst_entry)=200 no matter if CONFIG_NET_CLS_ROUTE=y/n. I tested many patches on my 16-core tigerton by moving tclassid to different place. It looks like tclassid could also have impact on performance. If moving tclassid before metrics, or just don't move tclassid, the performance isn't good. So I move it behind metrics. 2) Add comments before __refcnt. On 16-core tigerton: If CONFIG_NET_CLS_ROUTE=y, the result with below patch is about 18% better than the one without the patch; If CONFIG_NET_CLS_ROUTE=n, the result with below patch is about 30% better than the one without the patch. With 32bit 2.6.25-rc1 on 8-core stoakley, the new patch doesn't introduce regression. Thank Eric, Valdis, and David! Signed-off-by: Zhang Yanmin <yanmin.zhang@intel.com> Acked-by: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-03-13 05:52:37 +00:00
__u32 tclassid;
#else
__u32 __pad2;
[NET]: Fix tbench regression in 2.6.25-rc1 Comparing with kernel 2.6.24, tbench result has regression with 2.6.25-rc1. 1) On 2 quad-core processor stoakley: 4%. 2) On 4 quad-core processor tigerton: more than 30%. bisect located below patch. b4ce92775c2e7ff9cf79cca4e0a19c8c5fd6287b is first bad commit commit b4ce92775c2e7ff9cf79cca4e0a19c8c5fd6287b Author: Herbert Xu <herbert@gondor.apana.org.au> Date: Tue Nov 13 21:33:32 2007 -0800 [IPV6]: Move nfheader_len into rt6_info The dst member nfheader_len is only used by IPv6. It's also currently creating a rather ugly alignment hole in struct dst. Therefore this patch moves it from there into struct rt6_info. Above patch changes the cache line alignment, especially member __refcnt. I did a testing by adding 2 unsigned long pading before lastuse, so the 3 members, lastuse/__refcnt/__use, are moved to next cache line. The performance is recovered. I created a patch to rearrange the members in struct dst_entry. With Eric and Valdis Kletnieks's suggestion, I made finer arrangement. 1) Move tclassid under ops in case CONFIG_NET_CLS_ROUTE=y. So sizeof(dst_entry)=200 no matter if CONFIG_NET_CLS_ROUTE=y/n. I tested many patches on my 16-core tigerton by moving tclassid to different place. It looks like tclassid could also have impact on performance. If moving tclassid before metrics, or just don't move tclassid, the performance isn't good. So I move it behind metrics. 2) Add comments before __refcnt. On 16-core tigerton: If CONFIG_NET_CLS_ROUTE=y, the result with below patch is about 18% better than the one without the patch; If CONFIG_NET_CLS_ROUTE=n, the result with below patch is about 30% better than the one without the patch. With 32bit 2.6.25-rc1 on 8-core stoakley, the new patch doesn't introduce regression. Thank Eric, Valdis, and David! Signed-off-by: Zhang Yanmin <yanmin.zhang@intel.com> Acked-by: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-03-13 05:52:37 +00:00
#endif
/*
* Align __refcnt to a 64 bytes alignment
* (L1_CACHE_SIZE would be too much)
*/
#ifdef CONFIG_64BIT
long __pad_to_align_refcnt[2];
#endif
[NET]: Fix tbench regression in 2.6.25-rc1 Comparing with kernel 2.6.24, tbench result has regression with 2.6.25-rc1. 1) On 2 quad-core processor stoakley: 4%. 2) On 4 quad-core processor tigerton: more than 30%. bisect located below patch. b4ce92775c2e7ff9cf79cca4e0a19c8c5fd6287b is first bad commit commit b4ce92775c2e7ff9cf79cca4e0a19c8c5fd6287b Author: Herbert Xu <herbert@gondor.apana.org.au> Date: Tue Nov 13 21:33:32 2007 -0800 [IPV6]: Move nfheader_len into rt6_info The dst member nfheader_len is only used by IPv6. It's also currently creating a rather ugly alignment hole in struct dst. Therefore this patch moves it from there into struct rt6_info. Above patch changes the cache line alignment, especially member __refcnt. I did a testing by adding 2 unsigned long pading before lastuse, so the 3 members, lastuse/__refcnt/__use, are moved to next cache line. The performance is recovered. I created a patch to rearrange the members in struct dst_entry. With Eric and Valdis Kletnieks's suggestion, I made finer arrangement. 1) Move tclassid under ops in case CONFIG_NET_CLS_ROUTE=y. So sizeof(dst_entry)=200 no matter if CONFIG_NET_CLS_ROUTE=y/n. I tested many patches on my 16-core tigerton by moving tclassid to different place. It looks like tclassid could also have impact on performance. If moving tclassid before metrics, or just don't move tclassid, the performance isn't good. So I move it behind metrics. 2) Add comments before __refcnt. On 16-core tigerton: If CONFIG_NET_CLS_ROUTE=y, the result with below patch is about 18% better than the one without the patch; If CONFIG_NET_CLS_ROUTE=n, the result with below patch is about 30% better than the one without the patch. With 32bit 2.6.25-rc1 on 8-core stoakley, the new patch doesn't introduce regression. Thank Eric, Valdis, and David! Signed-off-by: Zhang Yanmin <yanmin.zhang@intel.com> Acked-by: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-03-13 05:52:37 +00:00
/*
* __refcnt wants to be on a different cache line from
* input/output/ops or performance tanks badly
*/
atomic_t __refcnt; /* client references */
int __use;
[NET]: Fix tbench regression in 2.6.25-rc1 Comparing with kernel 2.6.24, tbench result has regression with 2.6.25-rc1. 1) On 2 quad-core processor stoakley: 4%. 2) On 4 quad-core processor tigerton: more than 30%. bisect located below patch. b4ce92775c2e7ff9cf79cca4e0a19c8c5fd6287b is first bad commit commit b4ce92775c2e7ff9cf79cca4e0a19c8c5fd6287b Author: Herbert Xu <herbert@gondor.apana.org.au> Date: Tue Nov 13 21:33:32 2007 -0800 [IPV6]: Move nfheader_len into rt6_info The dst member nfheader_len is only used by IPv6. It's also currently creating a rather ugly alignment hole in struct dst. Therefore this patch moves it from there into struct rt6_info. Above patch changes the cache line alignment, especially member __refcnt. I did a testing by adding 2 unsigned long pading before lastuse, so the 3 members, lastuse/__refcnt/__use, are moved to next cache line. The performance is recovered. I created a patch to rearrange the members in struct dst_entry. With Eric and Valdis Kletnieks's suggestion, I made finer arrangement. 1) Move tclassid under ops in case CONFIG_NET_CLS_ROUTE=y. So sizeof(dst_entry)=200 no matter if CONFIG_NET_CLS_ROUTE=y/n. I tested many patches on my 16-core tigerton by moving tclassid to different place. It looks like tclassid could also have impact on performance. If moving tclassid before metrics, or just don't move tclassid, the performance isn't good. So I move it behind metrics. 2) Add comments before __refcnt. On 16-core tigerton: If CONFIG_NET_CLS_ROUTE=y, the result with below patch is about 18% better than the one without the patch; If CONFIG_NET_CLS_ROUTE=n, the result with below patch is about 30% better than the one without the patch. With 32bit 2.6.25-rc1 on 8-core stoakley, the new patch doesn't introduce regression. Thank Eric, Valdis, and David! Signed-off-by: Zhang Yanmin <yanmin.zhang@intel.com> Acked-by: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-03-13 05:52:37 +00:00
unsigned long lastuse;
union {
struct dst_entry *next;
struct rtable __rcu *rt_next;
struct rt6_info *rt6_next;
struct dn_route __rcu *dn_next;
};
};
net: Implement read-only protection and COW'ing of metrics. Routing metrics are now copy-on-write. Initially a route entry points it's metrics at a read-only location. If a routing table entry exists, it will point there. Else it will point at the all zero metric place-holder called 'dst_default_metrics'. The writeability state of the metrics is stored in the low bits of the metrics pointer, we have two bits left to spare if we want to store more states. For the initial implementation, COW is implemented simply via kmalloc. However future enhancements will change this to place the writable metrics somewhere else, in order to increase sharing. Very likely this "somewhere else" will be the inetpeer cache. Note also that this means that metrics updates may transiently fail if we cannot COW the metrics successfully. But even by itself, this patch should decrease memory usage and increase cache locality especially for routing workloads. In those cases the read-only metric copies stay in place and never get written to. TCP workloads where metrics get updated, and those rare cases where PMTU triggers occur, will take a very slight performance hit. But that hit will be alleviated when the long-term writable metrics move to a more sharable location. Since the metrics storage went from a u32 array of RTAX_MAX entries to what is essentially a pointer, some retooling of the dst_entry layout was necessary. Most importantly, we need to preserve the alignment of the reference count so that it doesn't share cache lines with the read-mostly state, as per Eric Dumazet's alignment assertion checks. The only non-trivial bit here is the move of the 'flags' member into the writeable cacheline. This is OK since we are always accessing the flags around the same moment when we made a modification to the reference count. Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-27 04:51:05 +00:00
extern u32 *dst_cow_metrics_generic(struct dst_entry *dst, unsigned long old);
extern const u32 dst_default_metrics[];
net: Implement read-only protection and COW'ing of metrics. Routing metrics are now copy-on-write. Initially a route entry points it's metrics at a read-only location. If a routing table entry exists, it will point there. Else it will point at the all zero metric place-holder called 'dst_default_metrics'. The writeability state of the metrics is stored in the low bits of the metrics pointer, we have two bits left to spare if we want to store more states. For the initial implementation, COW is implemented simply via kmalloc. However future enhancements will change this to place the writable metrics somewhere else, in order to increase sharing. Very likely this "somewhere else" will be the inetpeer cache. Note also that this means that metrics updates may transiently fail if we cannot COW the metrics successfully. But even by itself, this patch should decrease memory usage and increase cache locality especially for routing workloads. In those cases the read-only metric copies stay in place and never get written to. TCP workloads where metrics get updated, and those rare cases where PMTU triggers occur, will take a very slight performance hit. But that hit will be alleviated when the long-term writable metrics move to a more sharable location. Since the metrics storage went from a u32 array of RTAX_MAX entries to what is essentially a pointer, some retooling of the dst_entry layout was necessary. Most importantly, we need to preserve the alignment of the reference count so that it doesn't share cache lines with the read-mostly state, as per Eric Dumazet's alignment assertion checks. The only non-trivial bit here is the move of the 'flags' member into the writeable cacheline. This is OK since we are always accessing the flags around the same moment when we made a modification to the reference count. Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-27 04:51:05 +00:00
#define DST_METRICS_READ_ONLY 0x1UL
#define __DST_METRICS_PTR(Y) \
((u32 *)((Y) & ~DST_METRICS_READ_ONLY))
#define DST_METRICS_PTR(X) __DST_METRICS_PTR((X)->_metrics)
static inline bool dst_metrics_read_only(const struct dst_entry *dst)
{
return dst->_metrics & DST_METRICS_READ_ONLY;
}
extern void __dst_destroy_metrics_generic(struct dst_entry *dst, unsigned long old);
static inline void dst_destroy_metrics_generic(struct dst_entry *dst)
{
unsigned long val = dst->_metrics;
if (!(val & DST_METRICS_READ_ONLY))
__dst_destroy_metrics_generic(dst, val);
}
static inline u32 *dst_metrics_write_ptr(struct dst_entry *dst)
{
unsigned long p = dst->_metrics;
BUG_ON(!p);
net: Implement read-only protection and COW'ing of metrics. Routing metrics are now copy-on-write. Initially a route entry points it's metrics at a read-only location. If a routing table entry exists, it will point there. Else it will point at the all zero metric place-holder called 'dst_default_metrics'. The writeability state of the metrics is stored in the low bits of the metrics pointer, we have two bits left to spare if we want to store more states. For the initial implementation, COW is implemented simply via kmalloc. However future enhancements will change this to place the writable metrics somewhere else, in order to increase sharing. Very likely this "somewhere else" will be the inetpeer cache. Note also that this means that metrics updates may transiently fail if we cannot COW the metrics successfully. But even by itself, this patch should decrease memory usage and increase cache locality especially for routing workloads. In those cases the read-only metric copies stay in place and never get written to. TCP workloads where metrics get updated, and those rare cases where PMTU triggers occur, will take a very slight performance hit. But that hit will be alleviated when the long-term writable metrics move to a more sharable location. Since the metrics storage went from a u32 array of RTAX_MAX entries to what is essentially a pointer, some retooling of the dst_entry layout was necessary. Most importantly, we need to preserve the alignment of the reference count so that it doesn't share cache lines with the read-mostly state, as per Eric Dumazet's alignment assertion checks. The only non-trivial bit here is the move of the 'flags' member into the writeable cacheline. This is OK since we are always accessing the flags around the same moment when we made a modification to the reference count. Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-27 04:51:05 +00:00
if (p & DST_METRICS_READ_ONLY)
return dst->ops->cow_metrics(dst, p);
return __DST_METRICS_PTR(p);
}
/* This may only be invoked before the entry has reached global
* visibility.
*/
static inline void dst_init_metrics(struct dst_entry *dst,
const u32 *src_metrics,
bool read_only)
{
dst->_metrics = ((unsigned long) src_metrics) |
(read_only ? DST_METRICS_READ_ONLY : 0);
}
static inline void dst_copy_metrics(struct dst_entry *dest, const struct dst_entry *src)
{
u32 *dst_metrics = dst_metrics_write_ptr(dest);
if (dst_metrics) {
u32 *src_metrics = DST_METRICS_PTR(src);
memcpy(dst_metrics, src_metrics, RTAX_MAX * sizeof(u32));
}
}
static inline u32 *dst_metrics_ptr(struct dst_entry *dst)
{
return DST_METRICS_PTR(dst);
}
static inline u32
dst_metric_raw(const struct dst_entry *dst, const int metric)
{
net: Implement read-only protection and COW'ing of metrics. Routing metrics are now copy-on-write. Initially a route entry points it's metrics at a read-only location. If a routing table entry exists, it will point there. Else it will point at the all zero metric place-holder called 'dst_default_metrics'. The writeability state of the metrics is stored in the low bits of the metrics pointer, we have two bits left to spare if we want to store more states. For the initial implementation, COW is implemented simply via kmalloc. However future enhancements will change this to place the writable metrics somewhere else, in order to increase sharing. Very likely this "somewhere else" will be the inetpeer cache. Note also that this means that metrics updates may transiently fail if we cannot COW the metrics successfully. But even by itself, this patch should decrease memory usage and increase cache locality especially for routing workloads. In those cases the read-only metric copies stay in place and never get written to. TCP workloads where metrics get updated, and those rare cases where PMTU triggers occur, will take a very slight performance hit. But that hit will be alleviated when the long-term writable metrics move to a more sharable location. Since the metrics storage went from a u32 array of RTAX_MAX entries to what is essentially a pointer, some retooling of the dst_entry layout was necessary. Most importantly, we need to preserve the alignment of the reference count so that it doesn't share cache lines with the read-mostly state, as per Eric Dumazet's alignment assertion checks. The only non-trivial bit here is the move of the 'flags' member into the writeable cacheline. This is OK since we are always accessing the flags around the same moment when we made a modification to the reference count. Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-27 04:51:05 +00:00
u32 *p = DST_METRICS_PTR(dst);
return p[metric-1];
}
static inline u32
dst_metric(const struct dst_entry *dst, const int metric)
{
WARN_ON_ONCE(metric == RTAX_HOPLIMIT ||
metric == RTAX_ADVMSS ||
metric == RTAX_MTU);
return dst_metric_raw(dst, metric);
}
static inline u32
dst_metric_advmss(const struct dst_entry *dst)
{
u32 advmss = dst_metric_raw(dst, RTAX_ADVMSS);
if (!advmss)
advmss = dst->ops->default_advmss(dst);
return advmss;
}
static inline void dst_metric_set(struct dst_entry *dst, int metric, u32 val)
{
net: Implement read-only protection and COW'ing of metrics. Routing metrics are now copy-on-write. Initially a route entry points it's metrics at a read-only location. If a routing table entry exists, it will point there. Else it will point at the all zero metric place-holder called 'dst_default_metrics'. The writeability state of the metrics is stored in the low bits of the metrics pointer, we have two bits left to spare if we want to store more states. For the initial implementation, COW is implemented simply via kmalloc. However future enhancements will change this to place the writable metrics somewhere else, in order to increase sharing. Very likely this "somewhere else" will be the inetpeer cache. Note also that this means that metrics updates may transiently fail if we cannot COW the metrics successfully. But even by itself, this patch should decrease memory usage and increase cache locality especially for routing workloads. In those cases the read-only metric copies stay in place and never get written to. TCP workloads where metrics get updated, and those rare cases where PMTU triggers occur, will take a very slight performance hit. But that hit will be alleviated when the long-term writable metrics move to a more sharable location. Since the metrics storage went from a u32 array of RTAX_MAX entries to what is essentially a pointer, some retooling of the dst_entry layout was necessary. Most importantly, we need to preserve the alignment of the reference count so that it doesn't share cache lines with the read-mostly state, as per Eric Dumazet's alignment assertion checks. The only non-trivial bit here is the move of the 'flags' member into the writeable cacheline. This is OK since we are always accessing the flags around the same moment when we made a modification to the reference count. Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-27 04:51:05 +00:00
u32 *p = dst_metrics_write_ptr(dst);
net: Implement read-only protection and COW'ing of metrics. Routing metrics are now copy-on-write. Initially a route entry points it's metrics at a read-only location. If a routing table entry exists, it will point there. Else it will point at the all zero metric place-holder called 'dst_default_metrics'. The writeability state of the metrics is stored in the low bits of the metrics pointer, we have two bits left to spare if we want to store more states. For the initial implementation, COW is implemented simply via kmalloc. However future enhancements will change this to place the writable metrics somewhere else, in order to increase sharing. Very likely this "somewhere else" will be the inetpeer cache. Note also that this means that metrics updates may transiently fail if we cannot COW the metrics successfully. But even by itself, this patch should decrease memory usage and increase cache locality especially for routing workloads. In those cases the read-only metric copies stay in place and never get written to. TCP workloads where metrics get updated, and those rare cases where PMTU triggers occur, will take a very slight performance hit. But that hit will be alleviated when the long-term writable metrics move to a more sharable location. Since the metrics storage went from a u32 array of RTAX_MAX entries to what is essentially a pointer, some retooling of the dst_entry layout was necessary. Most importantly, we need to preserve the alignment of the reference count so that it doesn't share cache lines with the read-mostly state, as per Eric Dumazet's alignment assertion checks. The only non-trivial bit here is the move of the 'flags' member into the writeable cacheline. This is OK since we are always accessing the flags around the same moment when we made a modification to the reference count. Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-27 04:51:05 +00:00
if (p)
p[metric-1] = val;
}
static inline u32
dst_feature(const struct dst_entry *dst, u32 feature)
{
return dst_metric(dst, RTAX_FEATURES) & feature;
}
static inline u32 dst_mtu(const struct dst_entry *dst)
{
return dst->ops->mtu(dst);
}
/* RTT metrics are stored in milliseconds for user ABI, but used as jiffies */
static inline unsigned long dst_metric_rtt(const struct dst_entry *dst, int metric)
{
return msecs_to_jiffies(dst_metric(dst, metric));
}
static inline u32
dst_allfrag(const struct dst_entry *dst)
{
int ret = dst_feature(dst, RTAX_FEATURE_ALLFRAG);
return ret;
}
static inline int
dst_metric_locked(const struct dst_entry *dst, int metric)
{
return dst_metric(dst, RTAX_LOCK) & (1<<metric);
}
static inline void dst_hold(struct dst_entry *dst)
{
/*
* If your kernel compilation stops here, please check
* __pad_to_align_refcnt declaration in struct dst_entry
*/
BUILD_BUG_ON(offsetof(struct dst_entry, __refcnt) & 63);
atomic_inc(&dst->__refcnt);
}
static inline void dst_use(struct dst_entry *dst, unsigned long time)
{
dst_hold(dst);
dst->__use++;
dst->lastuse = time;
}
static inline void dst_use_noref(struct dst_entry *dst, unsigned long time)
{
dst->__use++;
dst->lastuse = time;
}
static inline struct dst_entry *dst_clone(struct dst_entry *dst)
{
if (dst)
atomic_inc(&dst->__refcnt);
return dst;
}
extern void dst_release(struct dst_entry *dst);
static inline void refdst_drop(unsigned long refdst)
{
if (!(refdst & SKB_DST_NOREF))
dst_release((struct dst_entry *)(refdst & SKB_DST_PTRMASK));
}
/**
* skb_dst_drop - drops skb dst
* @skb: buffer
*
* Drops dst reference count if a reference was taken.
*/
static inline void skb_dst_drop(struct sk_buff *skb)
{
if (skb->_skb_refdst) {
refdst_drop(skb->_skb_refdst);
skb->_skb_refdst = 0UL;
}
}
static inline void skb_dst_copy(struct sk_buff *nskb, const struct sk_buff *oskb)
{
nskb->_skb_refdst = oskb->_skb_refdst;
if (!(nskb->_skb_refdst & SKB_DST_NOREF))
dst_clone(skb_dst(nskb));
}
/**
* skb_dst_force - makes sure skb dst is refcounted
* @skb: buffer
*
* If dst is not yet refcounted, let's do it
*/
static inline void skb_dst_force(struct sk_buff *skb)
{
if (skb_dst_is_noref(skb)) {
WARN_ON(!rcu_read_lock_held());
skb->_skb_refdst &= ~SKB_DST_NOREF;
dst_clone(skb_dst(skb));
}
}
/**
* __skb_tunnel_rx - prepare skb for rx reinsert
* @skb: buffer
* @dev: tunnel device
* @net: netns for packet i/o
*
* After decapsulation, packet is going to re-enter (netif_rx()) our stack,
* so make some cleanups. (no accounting done)
*/
static inline void __skb_tunnel_rx(struct sk_buff *skb, struct net_device *dev,
struct net *net)
{
skb->dev = dev;
/*
* Clear rxhash so that we can recalulate the hash for the
* encapsulated packet, unless we have already determine the hash
* over the L4 4-tuple.
*/
if (!skb->l4_rxhash)
skb->rxhash = 0;
skb_set_queue_mapping(skb, 0);
skb_scrub_packet(skb, !net_eq(net, dev_net(dev)));
}
/**
* skb_tunnel_rx - prepare skb for rx reinsert
* @skb: buffer
* @dev: tunnel device
*
* After decapsulation, packet is going to re-enter (netif_rx()) our stack,
* so make some cleanups, and perform accounting.
* Note: this accounting is not SMP safe.
*/
static inline void skb_tunnel_rx(struct sk_buff *skb, struct net_device *dev,
struct net *net)
{
/* TODO : stats should be SMP safe */
dev->stats.rx_packets++;
dev->stats.rx_bytes += skb->len;
__skb_tunnel_rx(skb, dev, net);
}
/* Children define the path of the packet through the
* Linux networking. Thus, destinations are stackable.
*/
static inline struct dst_entry *skb_dst_pop(struct sk_buff *skb)
{
struct dst_entry *child = dst_clone(skb_dst(skb)->child);
skb_dst_drop(skb);
return child;
}
extern int dst_discard(struct sk_buff *skb);
extern void *dst_alloc(struct dst_ops *ops, struct net_device *dev,
int initial_ref, int initial_obsolete,
unsigned short flags);
extern void __dst_free(struct dst_entry *dst);
extern struct dst_entry *dst_destroy(struct dst_entry *dst);
static inline void dst_free(struct dst_entry *dst)
{
if (dst->obsolete > 0)
return;
if (!atomic_read(&dst->__refcnt)) {
dst = dst_destroy(dst);
if (!dst)
return;
}
__dst_free(dst);
}
static inline void dst_rcu_free(struct rcu_head *head)
{
struct dst_entry *dst = container_of(head, struct dst_entry, rcu_head);
dst_free(dst);
}
static inline void dst_confirm(struct dst_entry *dst)
{
dst->pending_confirm = 1;
}
static inline int dst_neigh_output(struct dst_entry *dst, struct neighbour *n,
struct sk_buff *skb)
{
const struct hh_cache *hh;
if (dst->pending_confirm) {
unsigned long now = jiffies;
dst->pending_confirm = 0;
/* avoid dirtying neighbour */
if (n->confirmed != now)
n->confirmed = now;
}
hh = &n->hh;
if ((n->nud_state & NUD_CONNECTED) && hh->hh_len)
return neigh_hh_output(hh, skb);
else
return n->output(n, skb);
}
static inline struct neighbour *dst_neigh_lookup(const struct dst_entry *dst, const void *daddr)
{
struct neighbour *n = dst->ops->neigh_lookup(dst, NULL, daddr);
return IS_ERR(n) ? NULL : n;
}
static inline struct neighbour *dst_neigh_lookup_skb(const struct dst_entry *dst,
struct sk_buff *skb)
{
struct neighbour *n = dst->ops->neigh_lookup(dst, skb, NULL);
return IS_ERR(n) ? NULL : n;
}
static inline void dst_link_failure(struct sk_buff *skb)
{
struct dst_entry *dst = skb_dst(skb);
if (dst && dst->ops && dst->ops->link_failure)
dst->ops->link_failure(skb);
}
static inline void dst_set_expires(struct dst_entry *dst, int timeout)
{
unsigned long expires = jiffies + timeout;
if (expires == 0)
expires = 1;
if (dst->expires == 0 || time_before(expires, dst->expires))
dst->expires = expires;
}
/* Output packet to network from transport. */
static inline int dst_output(struct sk_buff *skb)
{
return skb_dst(skb)->output(skb);
}
/* Input packet from network to transport. */
static inline int dst_input(struct sk_buff *skb)
{
return skb_dst(skb)->input(skb);
}
static inline struct dst_entry *dst_check(struct dst_entry *dst, u32 cookie)
{
if (dst->obsolete)
dst = dst->ops->check(dst, cookie);
return dst;
}
extern void dst_init(void);
/* Flags for xfrm_lookup flags argument. */
enum {
XFRM_LOOKUP_ICMP = 1 << 0,
};
struct flowi;
#ifndef CONFIG_XFRM
static inline struct dst_entry *xfrm_lookup(struct net *net,
struct dst_entry *dst_orig,
const struct flowi *fl, struct sock *sk,
int flags)
{
return dst_orig;
}
#else
extern struct dst_entry *xfrm_lookup(struct net *net, struct dst_entry *dst_orig,
const struct flowi *fl, struct sock *sk,
int flags);
#endif
#endif /* _NET_DST_H */