Fixes for net-2.6
Subsystems (3)
PATCH 1/1 Fixes for net-2.6
Hello,
I have two patches to fix XFRM. Can you check and apply them?
HEADLINES
---------
[XFRM] POLICY: Fix per-direction policy counter after flushing. [XFRM] STATE: Use destination address for src hash.
DIFFSTAT
--------
net/xfrm/xfrm_hash.h | 7 ++++--- net/xfrm/xfrm_policy.c | 4 ++-- net/xfrm/xfrm_state.c | 16 +++++++++------- 3 files changed, 15 insertions(+), 12 deletions(-)
CHANGESETS
----------
commit 90c1f7d3e1019b2885844b03088588268e38cec5
Author: Masahide NAKAMURA <[email protected]> Date: Sun Sep 24 14:46:59 2006 +0900
[XFRM] POLICY: Fix per-direction policy counter after flushing.
Currently when xfrm_policy_flush() is called per-direction policy counter is cleared. However flusing policy is performed for each type (i.e. main or sub) then it is not always true to make the counter zero.
net/xfrm/xfrm_policy.c +2 -2
@@ -787,6 +787,7 @@ void xfrm_policy_flush(u8 type)continue ;hlist_del (&pol->bydst);hlist_del (&pol->byidx); + xfrm_policy_count[dir]--;write_unlock_bh (&xfrm_policy_lock);xfrm_policy_kill (pol);
@@ -804,6 +805,7 @@ void xfrm_policy_flush(u8 type)continue ;hlist_del (&pol->bydst);hlist_del (&pol->byidx); + xfrm_policy_count[dir]--;write_unlock_bh (&xfrm_policy_lock);xfrm_policy_kill (pol);
@@ -812,8 +814,6 @@ void xfrm_policy_flush(u8 type) goto again2; } } - - xfrm_policy_count[dir] =0 ; }atomic_inc (&flow_cache_genid);write_unlock_bh (&xfrm_policy_lock);
---
commit e517421855d241f0b85a186b25e85d00eafa129f
Author: Masahide NAKAMURA <[email protected]> Date: Sat Sep 23 16:41:34 2006 +0900
[XFRM] STATE: Use destination address for src hash.
Src hash is introduced for Mobile IPv6 route optimization usage. On current kenrel code it is calculated with source address only. It results we uses the same hash value for outbound state (when the node has only one address for Mobile IPv6).
This patch use also destination address as peer information for src hash to be dispersed.
net/xfrm/xfrm_hash.h +4 -3
@@ -41,17 +41,18 @@ static inline unsigned int __xfrm_dst_hareturn (h ^ (h >>16 )) & hmask; } -static inline unsigned __xfrm_src_hash (xfrm_address_t *saddr, +static inline unsigned __xfrm_src_hash (xfrm_address_t *daddr, +xfrm_address_t *saddr,unsigned short family,unsigned int hmask) {unsigned int h = family;switch (family) {case AF_INET : - h ^=__xfrm4_addr_hash (saddr); + h ^=__xfrm4_daddr_saddr_hash (daddr, saddr);break ;case AF_INET6 : - h ^=__xfrm6_addr_hash (saddr); + h ^=__xfrm6_daddr_saddr_hash (daddr, saddr);break ; };return (h ^ (h >>16 )) & hmask;
net/xfrm/xfrm_state.c +9 -7
@@ -63,10 +63,11 @@ static inline unsigned int xfrm_dst_hashreturn __xfrm_dst_hash (daddr, saddr, reqid, family, xfrm_state_hmask); } -static inline unsigned int xfrm_src_hash (xfrm_address_t *addr, +static inline unsigned int xfrm_src_hash (xfrm_address_t *daddr, +xfrm_address_t *saddr,unsigned short family) { -return __xfrm_src_hash (addr, family, xfrm_state_hmask); +return __xfrm_src_hash (daddr, saddr, family, xfrm_state_hmask); } static inline unsigned int
@@ -92,7 +93,8 @@ static void xfrm_hash_transfer(struct hl nhashmask);hlist_add_head (&x->bydst, ndsttable+h); - h =__xfrm_src_hash (&x->props.saddr, x->props.family, + h =__xfrm_src_hash (&x->id.daddr, &x->props.saddr, + x->props.family, nhashmask);hlist_add_head (&x->bysrc, nsrctable+h);
@@ -458,7 +460,7 @@ static struct xfrm_state *__xfrm_state_lstatic struct xfrm_state *__xfrm_state_lookup_byaddr (xfrm_address_t *daddr,xfrm_address_t *saddr,u8 proto,unsigned short family) { -unsigned int h =xfrm_src_hash (saddr, family); +unsigned int h =xfrm_src_hash (daddr, saddr, family);struct xfrm_state *x;struct hlist_node *entry;
@@ -587,7 +589,7 @@ xfrm_state_find(xfrm_address_t *daddr, xif (km_query (x, tmpl, pol) ==0 ) { x->km.state =XFRM_STATE_ACQ ;hlist_add_head (&x->bydst, xfrm_state_bydst+h); - h =xfrm_src_hash (saddr, family); + h =xfrm_src_hash (daddr, saddr, family);hlist_add_head (&x->bysrc, xfrm_state_bysrc+h);if (x->id.spi) { h =xfrm_spi_hash (&x->id.daddr, x->id.spi, x->id.proto, family);
@@ -622,7 +624,7 @@ static void __xfrm_state_insert(struct x x->props.reqid, x->props.family);hlist_add_head (&x->bydst, xfrm_state_bydst+h); - h =xfrm_src_hash (&x->props.saddr, x->props.family); + h =xfrm_src_hash (&x->id.daddr, &x->props.saddr, x->props.family);hlist_add_head (&x->bysrc, xfrm_state_bysrc+h);if (x->id.spi) {
@@ -748,7 +750,7 @@ static struct xfrm_state *__find_acq_cor x->timer.expires = jiffies +XFRM_ACQ_EXPIRES *HZ ;add_timer (&x->timer);hlist_add_head (&x->bydst, xfrm_state_bydst+h); - h =xfrm_src_hash (saddr, family); + h =xfrm_src_hash (daddr, saddr, family);hlist_add_head (&x->bysrc, xfrm_state_bysrc+h);wake_up (&km_waitq); }
---
Signature · 1 line
Masahide NAKAMURA
David Miller
maintainerFrom: Masahide NAKAMURA <[email protected]>
Date: Tue, 03 Oct 2006 12:29:54 +0900
Quoted 8 lines
[XFRM] POLICY: Fix per-direction policy counter after flushing.
Currently when
xfrm_policy_flush()is called per-direction policy counter is cleared. However flusing policy is performed for each type (i.e. main or sub) then it is not always true to make the counter zero.Signed-off-by Masahide NAKAMURA <[email protected]>
The idea of this code is to avoid updating global state many many times during such a flush. This can be expensive and cause much SMP cacheline activity as other cpus read the counter in the routing lookup path.
I think what I'll do is reimplement this patch so that a local variable is used to maintain how many entries were removed, and then simply subtract that counter from xfrm_policy_count[dir] at the very end where the assignment to zero occurs.
Quoted 10 lines
[XFRM] STATE: Use destination address for src hash.
Src hash is introduced for Mobile IPv6 route optimization usage. On current kenrel code it is calculated with source address only. It results we uses the same hash value for outbound state (when the node has only one address for Mobile IPv6).
This patch use also destination address as peer information for src hash to be dispersed.Signed-off-by Masahide NAKAMURA <[email protected]>
This patch is fine, I will apply this, thank you.
Quoted 10 lines · David Miller
From: Masahide NAKAMURA <[email protected]>
Date: Tue, 03 Oct 2006 12:29:54 +0900Quoted 7 lines
[XFRM] POLICY: Fix per-direction policy counter after flushing.
Currently when
xfrm_policy_flush()is called per-direction policy counter is cleared. However flusing policy is performed for each type (i.e. main or sub) then it is not always true to make the counter zero.
Quoted 6 lines
Quoted 1 line
Signed-off-by Masahide NAKAMURA <[email protected]>The idea of this code is to avoid updating global state many many times during such a flush. This can be expensive and cause much SMP cacheline activity as other cpus read the counter in the routing lookup path.
Thanks for the clarify. My patch should have included such cacheline consideration.
Quoted 4 lines
I think what I'll do is reimplement this patch so that a local variable is used to maintain how many entries were removed, and then simply subtract that counter from xfrm_policy_count[dir] at the very end where the assignment to zero occurs.
I feel it's better idea now. I agree to apply it instead of my patch.
Regards,
Signature · 1 line
Masahide NAKAMURA