inetpeer: harden tree sizing and address side-channel eviction
Subsystems (3)
cover inetpeer: harden tree sizing and address side-channel eviction
This series addresses a vulnerability in the inetpeer subsystem where an off-path adversary can manipulate garbage collection to bypass IP-keyed ICMP rate limits, and hardens inetpeer memory usage under pressure.
The inetpeer rate limiting system stores peer entries in an RB-tree keyed deterministically by remote IP address. Because tree lookups walk the tree using lexicographical comparison (inetpeer_addr_cmp), an off-path attacker can predict the tree topology and the exact sequence of nodes traversed during lookups (the gc_stack candidate list). By combining predictable traversal with aggressive garbage collection triggered when the tree size exceeds inet_peer_threshold, an attacker can selectively force the eviction of targeted inet_peer nodes. When an evicted node is recreated on the next packet, its rate-limiting token bucket is reset to full capacity, creating a side-channel to bypass ICMP rate limits and probe for open UDP ports.
- Patch 1 enforces a hard limit on inet_peer allocations at 2x inet_peer_threshold to prevent unbounded slab memory growth if GC cannot free entries, and ensures inet_peer_xrlim_allow() fails closed (returns false) if peer allocation fails and returns NULL.
- Patch 2 randomizes RB-tree node ordering using SipHash with a secret boot-time key (inetpeer_hash_key), making tree layout and GC eviction paths unpredictable to off-path attackers.
Eric Dumazet (2):
inetpeer: enforce hard limit on tree size and fix NULL peer rate limit
inetpeer: randomize RB-tree node comparison using SipHash
include/net/inetpeer.h | 4 ++++ net/ipv4/inetpeer.c | 43 +++++++++++++++++++++++++++++++++++++----- 2 files changed, 42 insertions(+), 5 deletions(-)
Signature · 1 line
2.55.0.691.gc56d675ccc-goog
PATCH 1/2 inetpeer: enforce hard limit on tree size and fix NULL peer rate limit
Currently, inet_peer_threshold is used to trigger aggressive garbage collection when tree size reaches the threshold. However, if garbage collection is unable to reclaim candidate entries (e.g. due to held references or active fragment queues), base->total can grow without any upper bound, consuming excessive slab memory.
Fix this by enforcing a hard limit on inet_peer allocations when (u64)base->total reaches 2ULL * READ_ONCE(inet_peer_threshold). Casting to 64-bit unsigned avoids signed integer overflow if inet_peer_threshold is configured to large values via sysctl.
Additionally, when peer allocation fails and returns NULL, inet_peer_xrlim_allow() previously returned true, allowing packets without rate limiting. Fix this to return false when peer is NULL, ensuring rate limiting fails closed under memory pressure.
--- net/ipv4/inetpeer.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
net/ipv4/inetpeer.c +3 -2
@@ -192,7 +192,8 @@ struct inet_peer *inet_getpeer(struct inet_peer_base *base, gc_cnt =0 ; p =lookup (daddr, base, seq, gc_stack, &gc_cnt, &parent, &pp);if (!p) { - p =kmem_cache_alloc (peer_cachep,GFP_ATOMIC ); +if ((u64 )base->total <2ULL *READ_ONCE (inet_peer_threshold)) + p =kmem_cache_alloc (peer_cachep,GFP_ATOMIC );if (p) { p->daddr = *daddr; p->dtime = (__u32 )jiffies;
@@ -248,7 +249,7 @@ bool inet_peer_xrlim_allow(struct inet_peer *peer, int timeout)bool rc = false;if (!peer) -return true; +return false; token = otoken =READ_ONCE (peer->rate_tokens); now = jiffies;
Signature · 1 line
2.55.0.691.gc56d675ccc-goog
Re: [PATCH net 1/2] inetpeer: enforce hard limit on tree size and fix NULL peer rate limit
Quoted 17 lines · Eric Dumazet
Currently,
inet_peer_thresholdis used to trigger aggressive garbage collection when tree size reaches the threshold. However, if garbage collection is unable to reclaim candidate entries (e.g. due to held references or active fragment queues),base->totalcan grow without any upper bound, consuming excessive slab memory.Fix this by enforcing a hard limit on
inet_peerallocations when (u64)base->total reaches 2ULL *READ_ONCE(inet_peer_threshold). Casting to 64-bit unsigned avoids signed integer overflow ifinet_peer_thresholdis configured to large values via sysctl.Additionally, when peer allocation fails and returns NULL,
inet_peer_xrlim_allow()previously returned true, allowing packets without rate limiting. Fix this to return false when peer is NULL, ensuring rate limiting fails closed under memory pressure.
Quoted 4 lines
Fixes 1da177e4c3f4 ("Linux-2.6.12-rc2")Reported-by Michael Blunt <[email protected]>Signed-off-by Eric Dumazet <[email protected]>---
I will drop this patch, sashiko is not happy and I do not really care.
pw-bot: cr
PATCH 2/2 inetpeer: randomize RB-tree node comparison using SipHash
In reply to Eric DumazetThe inetpeer rate limiting system stores peer entries in a Red-Black tree keyed deterministically on the remote IP address. Because tree lookups walk the RB-tree using standard lexicographical comparisons (inetpeer_addr_cmp), an off-path adversary can predict the exact topology of the tree and the sequence of nodes traversed during lookups (the gc_stack candidate list).
By combining deterministic tree traversal with aggressive garbage collection (triggered when tree size exceeds inet_peer_threshold), an attacker can selectively force the eviction of targeted inet_peer nodes. When an evicted node is subsequently re-created upon receiving a new packet, its rate-limiting token bucket (rate_tokens, rate_last) is reset to full capacity. This creates a side-channel primitive allowing off-path attackers to bypass IP-keyed ICMP rate limits and infer open UDP ports (similar to SAD DNS style attacks).
Mitigate this by randomizing the RB-tree node comparison logic using SipHash with a secret boot-time key (inetpeer_hash_key). Nodes are ordered in the tree by SipHash(addr, key) rather than raw IP addresses. Because the secret key is unknown to external entities, the tree layout and lookup traversal paths are unpredictable to off-path adversaries, breaking the deterministic eviction gadget.
Cache the computed 64-bit SipHash (hash) in struct inet_peer and compute the target hash (dhash) once at the beginning of inet_getpeer() to avoid recomputing SipHash at every step of the RB-tree walk.
--- include/net/inetpeer.h | 4 ++++ net/ipv4/inetpeer.c | 38 +++++++++++++++++++++++++++++++++++--- 2 files changed, 39 insertions(+), 3 deletions(-)
include/net/inetpeer.h +4 -0
@@ -35,6 +35,7 @@ struct inetpeer_addr {struct inet_peer {struct rb_node rb_node; +u64 hash;struct inetpeer_addr daddr;u32 metrics[RTAX_MAX ];
@@ -125,6 +126,9 @@ static inline int inetpeer_addr_cmp(const struct inetpeer_addr *a, {int i, n; +if (a->family != b->family) +return a->family < b->family ?-1 :1 ; +if (a->family ==AF_INET ) n =sizeof (a->a4) /sizeof (u32);else
net/ipv4/inetpeer.c +35 -3
@@ -21,6 +21,7 @@#include <net/ip.h> #include <net/inetpeer.h> #include <net/secure_seq.h> +#include <linux/siphash.h> /* * Theory of operations.
@@ -52,6 +53,32 @@*/ static struct kmem_cache *peer_cachep__ro_after_init ; +static siphash_aligned_key_t inetpeer_hash_key__read_mostly ; + +static u64 inetpeer_addr_hash (const struct inetpeer_addr *a) +{ +if (a->family ==AF_INET ) +return siphash_2u32 ((__force u32 )a->a4.addr, a->a4.vif, + &inetpeer_hash_key); + +return siphash_4u32 ((__force u32 )a->a6.s6_addr32[0 ], + (__force u32 )a->a6.s6_addr32[1 ], + (__force u32 )a->a6.s6_addr32[2 ], + (__force u32 )a->a6.s6_addr32[3 ], + &inetpeer_hash_key); +} + +static int inetpeer_entry_cmp (u64 dhash, +const struct inetpeer_addr *daddr, +const struct inet_peer *p) +{ +if (dhash < p->hash) +return -1 ; +if (dhash > p->hash) +return 1 ; + +return inetpeer_addr_cmp (daddr, &p->daddr); +}void inet_peer_base_init (struct inet_peer_base *bp) {
@@ -73,6 +100,8 @@ void __init inet_initpeers(void) {u64 nr_entries; +get_random_bytes (&inetpeer_hash_key,sizeof (inetpeer_hash_key)); +/* 1% of physical memory */ nr_entries =div64_ul ((u64 )totalram_pages () <<PAGE_SHIFT ,100 *L1_CACHE_ALIGN (sizeof (struct inet_peer )));
@@ -84,6 +113,7 @@ void __init inet_initpeers(void)/* Called with rcu_read_lock() or base->lock held */ static struct inet_peer *lookup(const struct inetpeer_addr *daddr, + u64 dhash, struct inet_peer_base *base, unsigned int seq, struct inet_peer *gc_stack[],
@@ -105,7 +135,7 @@ static struct inet_peer *lookup(const struct inetpeer_addr *daddr,break ; parent = next; p = rb_entry(parent, struct inet_peer, rb_node); - cmp =inetpeer_addr_cmp (daddr, &p->daddr); + cmp =inetpeer_entry_cmp (dhash, daddr, p);if (cmp ==0 ) { now = jiffies;if (READ_ONCE (p->dtime) != now)
@@ -170,6 +200,7 @@ struct inet_peer *inet_getpeer(struct inet_peer_base *base, const struct inetpeer_addr *daddr) {struct inet_peer *p, *gc_stack[PEER_MAX_GC ]; +u64 dhash =inetpeer_addr_hash (daddr);struct rb_node **pp, *parent;unsigned int gc_cnt, seq;
@@ -177,7 +208,7 @@ struct inet_peer *inet_getpeer(struct inet_peer_base *base,* Because of a concurrent writer, we might not find an existing entry. */ seq =read_seqbegin (&base->lock); - p =lookup (daddr, base, seq,NULL , &gc_cnt, &parent, &pp); + p =lookup (daddr, dhash, base, seq,NULL , &gc_cnt, &parent, &pp);/* Make sure tree was not modified during our lookup. */ if (p && !read_seqretry (&base->lock, seq))
@@ -190,12 +221,13 @@ struct inet_peer *inet_getpeer(struct inet_peer_base *base,write_seqlock_bh (&base->lock); gc_cnt =0 ; - p =lookup (daddr, base, seq, gc_stack, &gc_cnt, &parent, &pp); + p =lookup (daddr, dhash, base, seq, gc_stack, &gc_cnt, &parent, &pp);if (!p) {if ((u64 )base->total <2ULL *READ_ONCE (inet_peer_threshold)) p =kmem_cache_alloc (peer_cachep,GFP_ATOMIC );if (p) { p->daddr = *daddr; + p->hash = dhash; p->dtime = (__u32 )jiffies;refcount_set (&p->refcnt,1 );atomic_set (&p->rid,0 );
Signature · 1 line
2.55.0.691.gc56d675ccc-goog
Re: [PATCH net 2/2] inetpeer: randomize RB-tree node comparison using SipHash
Quoted 26 lines · Eric Dumazet
The inetpeer rate limiting system stores peer entries in a Red-Black tree keyed deterministically on the remote IP address. Because tree lookups walk the RB-tree using standard lexicographical comparisons (
inetpeer_addr_cmp), an off-path adversary can predict the exact topology of the tree and the sequence of nodes traversed during lookups (thegc_stackcandidate list).By combining deterministic tree traversal with aggressive garbage collection (triggered when tree size exceeds
inet_peer_threshold), an attacker can selectively force the eviction of targetedinet_peernodes. When an evicted node is subsequently re-created upon receiving a new packet, its rate-limiting token bucket (rate_tokens,rate_last) is reset to full capacity. This creates a side-channel primitive allowing off-path attackers to bypass IP-keyed ICMP rate limits and infer open UDP ports (similar to SAD DNS style attacks).Mitigate this by randomizing the RB-tree node comparison logic using SipHash with a secret boot-time key (
inetpeer_hash_key). Nodes are ordered in the tree bySipHash(addr, key)rather than raw IP addresses. Because the secret key is unknown to external entities, the tree layout and lookup traversal paths are unpredictable to off-path adversaries, breaking the deterministic eviction gadget.Cache the computed 64-bit SipHash (hash) in struct
inet_peerand compute the target hash (dhash) once at the beginning ofinet_getpeer()to avoid recomputing SipHash at every step of the RB-tree walk.
Quoted 5 lines
Fixes b145425f269a ("inetpeer: remove AVL implementation in favor of RB tree")Reported-by Michael Blunt <[email protected]>Suggested-by Michael Blunt <[email protected]>Signed-off-by Eric Dumazet <[email protected]>---
I will send a V2 to address sashiko's feedback (
https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260812205946.356185-1-edumazet%40google.com
)
pw-bot: cr