xfrm: Add support for per cpu xfrm state handling.

Currently all flows for a certain SA must be processed by the same
cpu to avoid packet reordering and lock contention of the xfrm
state lock.

To get rid of this limitation, the IETF standardized per cpu SAs
in RFC 9611. This patch implements the xfrm part of it.

We add the cpu as a lookup key for xfrm states and a config option
to generate acquire messages for each cpu.

With that, we can have on each cpu a SA with identical traffic selector
so that flows can be processed in parallel on all cpus.

Signed-off-by: Steffen Klassert <[e-mail]>
Tested-by: Antony Antony <[e-mail]>
Tested-by: Tobias Brunner <[e-mail]>

include/net/xfrm.h

···
 	refcount_t		refcnt;
 	spinlock_t		lock;
 
+	u32			pcpu_num;
 	struct xfrm_id		id;
 	struct xfrm_selector	sel;
 	struct xfrm_mark	mark;
···
 	u32 spdhmcnt;
 };
 
-struct xfrm_state *xfrm_find_acq_byseq(struct net *net, u32 mark, u32 seq);
+struct xfrm_state *xfrm_find_acq_byseq(struct net *net, u32 mark, u32 seq, u32 pcpu_num);
 int xfrm_state_delete(struct xfrm_state *x);
 int xfrm_state_flush(struct net *net, u8 proto, bool task_valid, bool sync);
 int xfrm_dev_state_flush(struct net *net, struct net_device *dev, bool task_valid);
···
 int xfrm_alloc_spi(struct xfrm_state *x, u32 minspi, u32 maxspi,
 		   struct netlink_ext_ack *extack);
 struct xfrm_state *xfrm_find_acq(struct net *net, const struct xfrm_mark *mark,
-				 u8 mode, u32 reqid, u32 if_id, u8 proto,
+				 u8 mode, u32 reqid, u32 if_id, u32 pcpu_num, u8 proto,
 				 const xfrm_address_t *daddr,
 				 const xfrm_address_t *saddr, int create,
 				 unsigned short family);

include/uapi/linux/xfrm.h

···
 	XFRMA_MTIMER_THRESH,	/* __u32 in seconds for input SA */
 	XFRMA_SA_DIR,		/* __u8 */
 	XFRMA_NAT_KEEPALIVE_INTERVAL,	/* __u32 in seconds for NAT keepalive */
+	XFRMA_SA_PCPU,		/* __u32 */
 	__XFRMA_MAX
 
 #define XFRMA_OUTPUT_MARK XFRMA_SET_MARK	/* Compatibility */
···
 #define XFRM_POLICY_LOCALOK	1	/* Allow user to override global policy */
 	/* Automatically expand selector to include matching ICMP payloads. */
 #define XFRM_POLICY_ICMP	2
+#define XFRM_POLICY_CPU_ACQUIRE	4
 	__u8				share;
 };
 

net/key/af_key.c

···
 	}
 
 	if (hdr->sadb_msg_seq) {
-		x = xfrm_find_acq_byseq(net, DUMMY_MARK, hdr->sadb_msg_seq);
+		x = xfrm_find_acq_byseq(net, DUMMY_MARK, hdr->sadb_msg_seq, UINT_MAX);
 		if (x && !xfrm_addr_equal(&x->id.daddr, xdaddr, family)) {
 			xfrm_state_put(x);
 			x = NULL;
···
 	}
 
 	if (!x)
-		x = xfrm_find_acq(net, &dummy_mark, mode, reqid, 0, proto, xdaddr, xsaddr, 1, family);
+		x = xfrm_find_acq(net, &dummy_mark, mode, reqid, 0, UINT_MAX,
+				  proto, xdaddr, xsaddr, 1, family);
 
 	if (x == NULL)
 		return -ENOENT;
···
 	if (hdr->sadb_msg_seq == 0 || hdr->sadb_msg_errno == 0)
 		return 0;
 
-	x = xfrm_find_acq_byseq(net, DUMMY_MARK, hdr->sadb_msg_seq);
+	x = xfrm_find_acq_byseq(net, DUMMY_MARK, hdr->sadb_msg_seq, UINT_MAX);
 	if (x == NULL)
 		return 0;
 

net/xfrm/xfrm_compat.c

···
 	[XFRMA_MTIMER_THRESH]	= { .type = NLA_U32 },
 	[XFRMA_SA_DIR]          = NLA_POLICY_RANGE(NLA_U8, XFRM_SA_DIR_IN, XFRM_SA_DIR_OUT),
 	[XFRMA_NAT_KEEPALIVE_INTERVAL]	= { .type = NLA_U32 },
+	[XFRMA_SA_PCPU]		= { .type = NLA_U32 },
 };
 
 static struct nlmsghdr *xfrm_nlmsg_put_compat(struct sk_buff *skb,
···
 	case XFRMA_MTIMER_THRESH:
 	case XFRMA_SA_DIR:
 	case XFRMA_NAT_KEEPALIVE_INTERVAL:
+	case XFRMA_SA_PCPU:
 		return xfrm_nla_cpy(dst, src, nla_len(src));
 	default:
-		BUILD_BUG_ON(XFRMA_MAX != XFRMA_NAT_KEEPALIVE_INTERVAL);
+		BUILD_BUG_ON(XFRMA_MAX != XFRMA_SA_PCPU);
 		pr_warn_once("unsupported nla_type %d\n", src->nla_type);
 		return -EOPNOTSUPP;
 	}
···
 	int err;
 
 	if (type > XFRMA_MAX) {
-		BUILD_BUG_ON(XFRMA_MAX != XFRMA_NAT_KEEPALIVE_INTERVAL);
+		BUILD_BUG_ON(XFRMA_MAX != XFRMA_SA_PCPU);
 		NL_SET_ERR_MSG(extack, "Bad attribute");
 		return -EOPNOTSUPP;
 	}

net/xfrm/xfrm_state.c

···
 		x->lft.hard_packet_limit = XFRM_INF;
 		x->replay_maxage = 0;
 		x->replay_maxdiff = 0;
+		x->pcpu_num = UINT_MAX;
 		spin_lock_init(&x->lock);
 	}
 	return x;
···
 			       struct xfrm_state **best, int *acq_in_progress,
 			       int *error)
 {
+	/* We need the cpu id just as a lookup key,
+	 * we don't require it to be stable.
+	 */
+	unsigned int pcpu_id = get_cpu();
+	put_cpu();
+
 	/* Resolution logic:
 	 * 1. There is a valid state with matching selector. Done.
 	 * 2. Valid state with inappropriate selector. Skip.
···
 							&fl->u.__fl_common))
 			return;
 
+		if (x->pcpu_num != UINT_MAX && x->pcpu_num != pcpu_id)
+			return;
+
 		if (!*best ||
+		    ((*best)->pcpu_num == UINT_MAX && x->pcpu_num == pcpu_id) ||
 		    (*best)->km.dying > x->km.dying ||
 		    ((*best)->km.dying == x->km.dying &&
 		     (*best)->curlft.add_time < x->curlft.add_time))
 			*best = x;
 	} else if (x->km.state == XFRM_STATE_ACQ) {
-		*acq_in_progress = 1;
+		if (!*best || x->pcpu_num == pcpu_id)
+			*acq_in_progress = 1;
 	} else if (x->km.state == XFRM_STATE_ERROR ||
 		   x->km.state == XFRM_STATE_EXPIRED) {
 		if ((!x->sel.family ||
···
 	unsigned short encap_family = tmpl->encap_family;
 	unsigned int sequence;
 	struct km_event c;
+	unsigned int pcpu_id;
+
+	/* We need the cpu id just as a lookup key,
+	 * we don't require it to be stable.
+	 */
+	pcpu_id = get_cpu();
+	put_cpu();
 
 	to_put = NULL;
 
···
 	}
 
 found:
-	x = best;
+	if (!(pol->flags & XFRM_POLICY_CPU_ACQUIRE) ||
+	    (best && (best->pcpu_num == pcpu_id)))
+		x = best;
+
 	if (!x && !error && !acquire_in_progress) {
 		if (tmpl->id.spi &&
 		    (x0 = __xfrm_state_lookup_all(net, mark, daddr,
···
 		xfrm_init_tempstate(x, fl, tmpl, daddr, saddr, family);
 		memcpy(&x->mark, &pol->mark, sizeof(x->mark));
 		x->if_id = if_id;
+		if ((pol->flags & XFRM_POLICY_CPU_ACQUIRE) && best)
+			x->pcpu_num = pcpu_id;
 
 		error = security_xfrm_state_alloc_acquire(x, pol->security, fl->flowi_secid);
 		if (error) {
···
 			x = NULL;
 			error = -ESRCH;
 		}
+
+		/* Use the already installed 'fallback' while the CPU-specific
+		 * SA acquire is handled*/
+		if (best)
+			x = best;
 	}
 out:
 	if (x) {
···
 	unsigned int h;
 	u32 mark = xnew->mark.v & xnew->mark.m;
 	u32 if_id = xnew->if_id;
+	u32 cpu_id = xnew->pcpu_num;
 
 	h = xfrm_dst_hash(net, &xnew->id.daddr, &xnew->props.saddr, reqid, family);
 	hlist_for_each_entry(x, net->xfrm.state_bydst+h, bydst) {
 		if (x->props.family	== family &&
 		    x->props.reqid	== reqid &&
 		    x->if_id		== if_id &&
+		    x->pcpu_num		== cpu_id &&
 		    (mark & x->mark.m) == x->mark.v &&
 		    xfrm_addr_equal(&x->id.daddr, &xnew->id.daddr, family) &&
 		    xfrm_addr_equal(&x->props.saddr, &xnew->props.saddr, family))
···
 static struct xfrm_state *__find_acq_core(struct net *net,
 					  const struct xfrm_mark *m,
 					  unsigned short family, u8 mode,
-					  u32 reqid, u32 if_id, u8 proto,
+					  u32 reqid, u32 if_id, u32 pcpu_num, u8 proto,
 					  const xfrm_address_t *daddr,
 					  const xfrm_address_t *saddr,
 					  int create)
···
 		    x->id.spi       != 0 ||
 		    x->id.proto	    != proto ||
 		    (mark & x->mark.m) != x->mark.v ||
+		    x->pcpu_num != pcpu_num ||
 		    !xfrm_addr_equal(&x->id.daddr, daddr, family) ||
 		    !xfrm_addr_equal(&x->props.saddr, saddr, family))
 			continue;
···
 			break;
 		}
 
+		x->pcpu_num = pcpu_num;
 		x->km.state = XFRM_STATE_ACQ;
 		x->id.proto = proto;
 		x->props.family = family;
···
 	return x;
 }
 
-static struct xfrm_state *__xfrm_find_acq_byseq(struct net *net, u32 mark, u32 seq);
+static struct xfrm_state *__xfrm_find_acq_byseq(struct net *net, u32 mark, u32 seq, u32 pcpu_num);
 
 int xfrm_state_add(struct xfrm_state *x)
 {
···
 	}
 
 	if (use_spi && x->km.seq) {
-		x1 = __xfrm_find_acq_byseq(net, mark, x->km.seq);
+		x1 = __xfrm_find_acq_byseq(net, mark, x->km.seq, x->pcpu_num);
 		if (x1 && ((x1->id.proto != x->id.proto) ||
 		    !xfrm_addr_equal(&x1->id.daddr, &x->id.daddr, family))) {
 			to_put = x1;
···
 
 	if (use_spi && !x1)
 		x1 = __find_acq_core(net, &x->mark, family, x->props.mode,
-				     x->props.reqid, x->if_id, x->id.proto,
+				     x->props.reqid, x->if_id, x->pcpu_num, x->id.proto,
 				     &x->id.daddr, &x->props.saddr, 0);
 
 	__xfrm_state_bump_genids(x);
···
 	x->props.flags = orig->props.flags;
 	x->props.extra_flags = orig->props.extra_flags;
 
+	x->pcpu_num = orig->pcpu_num;
 	x->if_id = orig->if_id;
 	x->tfcpad = orig->tfcpad;
 	x->replay_maxdiff = orig->replay_maxdiff;
···
 
 struct xfrm_state *
 xfrm_find_acq(struct net *net, const struct xfrm_mark *mark, u8 mode, u32 reqid,
-	      u32 if_id, u8 proto, const xfrm_address_t *daddr,
+	      u32 if_id, u32 pcpu_num, u8 proto, const xfrm_address_t *daddr,
 	      const xfrm_address_t *saddr, int create, unsigned short family)
 {
 	struct xfrm_state *x;
 
 	spin_lock_bh(&net->xfrm.xfrm_state_lock);
-	x = __find_acq_core(net, mark, family, mode, reqid, if_id, proto, daddr, saddr, create);
+	x = __find_acq_core(net, mark, family, mode, reqid, if_id, pcpu_num,
+			    proto, daddr, saddr, create);
 	spin_unlock_bh(&net->xfrm.xfrm_state_lock);
 
 	return x;
···
 
 /* Silly enough, but I'm lazy to build resolution list */
 
-static struct xfrm_state *__xfrm_find_acq_byseq(struct net *net, u32 mark, u32 seq)
+static struct xfrm_state *__xfrm_find_acq_byseq(struct net *net, u32 mark, u32 seq, u32 pcpu_num)
 {
 	unsigned int h = xfrm_seq_hash(net, seq);
 	struct xfrm_state *x;
···
 	hlist_for_each_entry_rcu(x, net->xfrm.state_byseq + h, byseq) {
 		if (x->km.seq == seq &&
 		    (mark & x->mark.m) == x->mark.v &&
+		    x->pcpu_num == pcpu_num &&
 		    x->km.state == XFRM_STATE_ACQ) {
 			xfrm_state_hold(x);
 			return x;
···
 	return NULL;
 }
 
-struct xfrm_state *xfrm_find_acq_byseq(struct net *net, u32 mark, u32 seq)
+struct xfrm_state *xfrm_find_acq_byseq(struct net *net, u32 mark, u32 seq, u32 pcpu_num)
 {
 	struct xfrm_state *x;
 
 	spin_lock_bh(&net->xfrm.xfrm_state_lock);
-	x = __xfrm_find_acq_byseq(net, mark, seq);
+	x = __xfrm_find_acq_byseq(net, mark, seq, pcpu_num);
 	spin_unlock_bh(&net->xfrm.xfrm_state_lock);
 	return x;
 }

net/xfrm/xfrm_user.c

···
 		}
 	}
 
+	if (!sa_dir && attrs[XFRMA_SA_PCPU]) {
+		NL_SET_ERR_MSG(extack, "SA_PCPU only supported with SA_DIR");
+		err = -EINVAL;
+		goto out;
+	}
+
 out:
 	return err;
 }
···
 	if (attrs[XFRMA_NAT_KEEPALIVE_INTERVAL])
 		x->nat_keepalive_interval =
 			nla_get_u32(attrs[XFRMA_NAT_KEEPALIVE_INTERVAL]);
+
+	if (attrs[XFRMA_SA_PCPU]) {
+		x->pcpu_num = nla_get_u32(attrs[XFRMA_SA_PCPU]);
+		if (x->pcpu_num >= num_possible_cpus())
+			goto error;
+	}
 
 	err = __xfrm_init_state(x, false, attrs[XFRMA_OFFLOAD_DEV], extack);
 	if (err)
···
 		if (ret)
 			goto out;
 	}
+	if (x->pcpu_num != UINT_MAX) {
+		ret = nla_put_u32(skb, XFRMA_SA_PCPU, x->pcpu_num);
+		if (ret)
+			goto out;
+	}
 	if (x->dir)
 		ret = nla_put_u8(skb, XFRMA_SA_DIR, x->dir);
 
···
 	u32 mark;
 	struct xfrm_mark m;
 	u32 if_id = 0;
+	u32 pcpu_num = UINT_MAX;
 
 	p = nlmsg_data(nlh);
 	err = verify_spi_info(p->info.id.proto, p->min, p->max, extack);
···
 	if (attrs[XFRMA_IF_ID])
 		if_id = nla_get_u32(attrs[XFRMA_IF_ID]);
 
+	if (attrs[XFRMA_SA_PCPU]) {
+		pcpu_num = nla_get_u32(attrs[XFRMA_SA_PCPU]);
+		if (pcpu_num >= num_possible_cpus()) {
+			err = -EINVAL;
+			goto out_noput;
+		}
+	}
+
 	if (p->info.seq) {
-		x = xfrm_find_acq_byseq(net, mark, p->info.seq);
+		x = xfrm_find_acq_byseq(net, mark, p->info.seq, pcpu_num);
 		if (x && !xfrm_addr_equal(&x->id.daddr, daddr, family)) {
 			xfrm_state_put(x);
 			x = NULL;
···
 
 	if (!x)
 		x = xfrm_find_acq(net, &m, p->info.mode, p->info.reqid,
-				  if_id, p->info.id.proto, daddr,
+				  if_id, pcpu_num, p->info.id.proto, daddr,
 				  &p->info.saddr, 1,
 				  family);
 	err = -ENOENT;
···
 	       + nla_total_size(sizeof(struct xfrm_mark))
 	       + nla_total_size(4) /* XFRM_AE_RTHR */
 	       + nla_total_size(4) /* XFRM_AE_ETHR */
-	       + nla_total_size(sizeof(x->dir)); /* XFRMA_SA_DIR */
+	       + nla_total_size(sizeof(x->dir)) /* XFRMA_SA_DIR */
+	       + nla_total_size(4); /* XFRMA_SA_PCPU */
 }
 
 static int build_aevent(struct sk_buff *skb, struct xfrm_state *x, const struct km_event *c)
···
 	err = xfrm_if_id_put(skb, x->if_id);
 	if (err)
 		goto out_cancel;
+	if (x->pcpu_num != UINT_MAX)
+		err = nla_put_u32(skb, XFRMA_SA_PCPU, x->pcpu_num);
 
 	if (x->dir) {
 		err = nla_put_u8(skb, XFRMA_SA_DIR, x->dir);
···
 		goto nomem;
 
 	xfrm_mark_get(attrs, &mark);
+
+	if (attrs[XFRMA_SA_PCPU]) {
+		x->pcpu_num = nla_get_u32(attrs[XFRMA_SA_PCPU]);
+		err = -EINVAL;
+		if (x->pcpu_num >= num_possible_cpus())
+			goto free_state;
+	}
 
 	err = verify_newpolicy_info(&ua->policy, extack);
 	if (err)
···
 	[XFRMA_MTIMER_THRESH]   = { .type = NLA_U32 },
 	[XFRMA_SA_DIR]          = NLA_POLICY_RANGE(NLA_U8, XFRM_SA_DIR_IN, XFRM_SA_DIR_OUT),
 	[XFRMA_NAT_KEEPALIVE_INTERVAL] = { .type = NLA_U32 },
+	[XFRMA_SA_PCPU]		= { .type = NLA_U32 },
 };
 EXPORT_SYMBOL_GPL(xfrma_policy);
 
···
 {
 	return NLMSG_ALIGN(sizeof(struct xfrm_user_expire)) +
 	       nla_total_size(sizeof(struct xfrm_mark)) +
-	       nla_total_size(sizeof_field(struct xfrm_state, dir));
+	       nla_total_size(sizeof_field(struct xfrm_state, dir)) +
+	       nla_total_size(4); /* XFRMA_SA_PCPU */
 }
 
 static int build_expire(struct sk_buff *skb, struct xfrm_state *x, const struct km_event *c)
···
 	err = xfrm_if_id_put(skb, x->if_id);
 	if (err)
 		return err;
+	if (x->pcpu_num != UINT_MAX) {
+		err = nla_put_u32(skb, XFRMA_SA_PCPU, x->pcpu_num);
+		if (err)
+			return err;
+	}
 
 	if (x->dir) {
 		err = nla_put_u8(skb, XFRMA_SA_DIR, x->dir);
···
 	}
 	if (x->if_id)
 		l += nla_total_size(sizeof(x->if_id));
+	if (x->pcpu_num)
+		l += nla_total_size(sizeof(x->pcpu_num));
 
 	/* Must count x->lastused as it may become non-zero behind our back. */
 	l += nla_total_size_64bit(sizeof(u64));
···
 	       + nla_total_size(sizeof(struct xfrm_user_tmpl) * xp->xfrm_nr)
 	       + nla_total_size(sizeof(struct xfrm_mark))
 	       + nla_total_size(xfrm_user_sec_ctx_size(x->security))
+	       + nla_total_size(4) /* XFRMA_SA_PCPU */
 	       + userpolicy_type_attrsize();
 }
 
···
 		err = xfrm_if_id_put(skb, xp->if_id);
 	if (!err && xp->xdo.dev)
 		err = copy_user_offload(&xp->xdo, skb);
+	if (!err && x->pcpu_num != UINT_MAX)
+		err = nla_put_u32(skb, XFRMA_SA_PCPU, x->pcpu_num);
 	if (err) {
 		nlmsg_cancel(skb, nlh);
 		return err;