net/tcp: Prepare tcp_md5sig_pool for TCP-AO
authorDmitry Safonov <dima@arista.com>
Mon, 23 Oct 2023 19:21:53 +0000 (20:21 +0100)
committerDavid S. Miller <davem@davemloft.net>
Fri, 27 Oct 2023 09:35:44 +0000 (10:35 +0100)
TCP-AO, similarly to TCP-MD5, needs to allocate tfms on a slow-path,
which is setsockopt() and use crypto ahash requests on fast paths,
which are RX/TX softirqs. Also, it needs a temporary/scratch buffer
for preparing the hash.

Rework tcp_md5sig_pool in order to support other hashing algorithms
than MD5. It will make it possible to share pre-allocated crypto_ahash
descriptors and scratch area between all TCP hash users.

Internally tcp_sigpool calls crypto_clone_ahash() API over pre-allocated
crypto ahash tfm. Kudos to Herbert, who provided this new crypto API.

I was a little concerned over GFP_ATOMIC allocations of ahash and
crypto_request in RX/TX (see tcp_sigpool_start()), so I benchmarked both
"backends" with different algorithms, using patched version of iperf3[2].
On my laptop with i7-7600U @ 2.80GHz:

                         clone-tfm                per-CPU-requests
TCP-MD5                  2.25 Gbits/sec           2.30 Gbits/sec
TCP-AO(hmac(sha1))       2.53 Gbits/sec           2.54 Gbits/sec
TCP-AO(hmac(sha512))     1.67 Gbits/sec           1.64 Gbits/sec
TCP-AO(hmac(sha384))     1.77 Gbits/sec           1.80 Gbits/sec
TCP-AO(hmac(sha224))     1.29 Gbits/sec           1.30 Gbits/sec
TCP-AO(hmac(sha3-512))    481 Mbits/sec            480 Mbits/sec
TCP-AO(hmac(md5))        2.07 Gbits/sec           2.12 Gbits/sec
TCP-AO(hmac(rmd160))     1.01 Gbits/sec            995 Mbits/sec
TCP-AO(cmac(aes128))     [not supporetd yet]      2.11 Gbits/sec

So, it seems that my concerns don't have strong grounds and per-CPU
crypto_request allocation can be dropped/removed from tcp_sigpool once
ciphers get crypto_clone_ahash() support.

[1]: https://lore.kernel.org/all/ZDefxOq6Ax0JeTRH@gondor.apana.org.au/T/#u
[2]: https://github.com/0x7f454c46/iperf/tree/tcp-md5-ao
Signed-off-by: Dmitry Safonov <dima@arista.com>
Reviewed-by: Steen Hegelund <Steen.Hegelund@microchip.com>
Acked-by: David Ahern <dsahern@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/net/tcp.h
net/ipv4/Kconfig
net/ipv4/Makefile
net/ipv4/tcp.c
net/ipv4/tcp_ipv4.c
net/ipv4/tcp_minisocks.c
net/ipv4/tcp_sigpool.c [new file with mode: 0644]
net/ipv6/tcp_ipv6.c

index 993b7fcd4e461fd8b0d4fa7da5434d8fcd319fdc..f6e2db5292b50e48425845906a04231a8899ae38 100644 (file)
@@ -1737,12 +1737,39 @@ union tcp_md5sum_block {
 #endif
 };
 
-/* - pool: digest algorithm, hash description and scratch buffer */
-struct tcp_md5sig_pool {
-       struct ahash_request    *md5_req;
-       void                    *scratch;
+/*
+ * struct tcp_sigpool - per-CPU pool of ahash_requests
+ * @scratch: per-CPU temporary area, that can be used between
+ *          tcp_sigpool_start() and tcp_sigpool_end() to perform
+ *          crypto request
+ * @req: pre-allocated ahash request
+ */
+struct tcp_sigpool {
+       void *scratch;
+       struct ahash_request *req;
 };
 
+int tcp_sigpool_alloc_ahash(const char *alg, size_t scratch_size);
+void tcp_sigpool_get(unsigned int id);
+void tcp_sigpool_release(unsigned int id);
+int tcp_sigpool_hash_skb_data(struct tcp_sigpool *hp,
+                             const struct sk_buff *skb,
+                             unsigned int header_len);
+
+/**
+ * tcp_sigpool_start - disable bh and start using tcp_sigpool_ahash
+ * @id: tcp_sigpool that was previously allocated by tcp_sigpool_alloc_ahash()
+ * @c: returned tcp_sigpool for usage (uninitialized on failure)
+ *
+ * Returns 0 on success, error otherwise.
+ */
+int tcp_sigpool_start(unsigned int id, struct tcp_sigpool *c);
+/**
+ * tcp_sigpool_end - enable bh and stop using tcp_sigpool
+ * @c: tcp_sigpool context that was returned by tcp_sigpool_start()
+ */
+void tcp_sigpool_end(struct tcp_sigpool *c);
+size_t tcp_sigpool_algo(unsigned int id, char *buf, size_t buf_len);
 /* - functions */
 int tcp_v4_md5_hash_skb(char *md5_hash, const struct tcp_md5sig_key *key,
                        const struct sock *sk, const struct sk_buff *skb);
@@ -1798,17 +1825,12 @@ tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
 #define tcp_twsk_md5_key(twsk) NULL
 #endif
 
-bool tcp_alloc_md5sig_pool(void);
-
-struct tcp_md5sig_pool *tcp_get_md5sig_pool(void);
-static inline void tcp_put_md5sig_pool(void)
-{
-       local_bh_enable();
-}
+int tcp_md5_alloc_sigpool(void);
+void tcp_md5_release_sigpool(void);
+void tcp_md5_add_sigpool(void);
+extern int tcp_md5_sigpool_id;
 
-int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *, const struct sk_buff *,
-                         unsigned int header_len);
-int tcp_md5_hash_key(struct tcp_md5sig_pool *hp,
+int tcp_md5_hash_key(struct tcp_sigpool *hp,
                     const struct tcp_md5sig_key *key);
 
 /* From tcp_fastopen.c */
index 2dfb12230f08943e52c39e36e22616db794cd1f7..89e2ab023272fa06b06e645352feac627ea26242 100644 (file)
@@ -741,10 +741,14 @@ config DEFAULT_TCP_CONG
        default "bbr" if DEFAULT_BBR
        default "cubic"
 
+config TCP_SIGPOOL
+       tristate
+
 config TCP_MD5SIG
        bool "TCP: MD5 Signature Option support (RFC2385)"
        select CRYPTO
        select CRYPTO_MD5
+       select TCP_SIGPOOL
        help
          RFC2385 specifies a method of giving MD5 protection to TCP sessions.
          Its main (only?) use is to protect BGP sessions between core routers
index b18ba8ef93ad2b6c292c2bdf0c47ac22bc74127a..cd760793cfcbae5862e51f2d5308309b1edafcac 100644 (file)
@@ -62,6 +62,7 @@ obj-$(CONFIG_TCP_CONG_SCALABLE) += tcp_scalable.o
 obj-$(CONFIG_TCP_CONG_LP) += tcp_lp.o
 obj-$(CONFIG_TCP_CONG_YEAH) += tcp_yeah.o
 obj-$(CONFIG_TCP_CONG_ILLINOIS) += tcp_illinois.o
+obj-$(CONFIG_TCP_SIGPOOL) += tcp_sigpool.o
 obj-$(CONFIG_NET_SOCK_MSG) += tcp_bpf.o
 obj-$(CONFIG_BPF_SYSCALL) += udp_bpf.o
 obj-$(CONFIG_NETLABEL) += cipso_ipv4.o
index 156264531124287d5212c9d9d28efcd4c3f6a184..dca9ca2f10819063b74389895041e4214fafeaea 100644 (file)
@@ -4305,141 +4305,52 @@ int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
 EXPORT_SYMBOL(tcp_getsockopt);
 
 #ifdef CONFIG_TCP_MD5SIG
-static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
-static DEFINE_MUTEX(tcp_md5sig_mutex);
-static bool tcp_md5sig_pool_populated = false;
+int tcp_md5_sigpool_id = -1;
+EXPORT_SYMBOL_GPL(tcp_md5_sigpool_id);
 
-static void __tcp_alloc_md5sig_pool(void)
+int tcp_md5_alloc_sigpool(void)
 {
-       struct crypto_ahash *hash;
-       int cpu;
-
-       hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
-       if (IS_ERR(hash))
-               return;
-
-       for_each_possible_cpu(cpu) {
-               void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
-               struct ahash_request *req;
-
-               if (!scratch) {
-                       scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
-                                              sizeof(struct tcphdr),
-                                              GFP_KERNEL,
-                                              cpu_to_node(cpu));
-                       if (!scratch)
-                               return;
-                       per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
-               }
-               if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
-                       continue;
-
-               req = ahash_request_alloc(hash, GFP_KERNEL);
-               if (!req)
-                       return;
-
-               ahash_request_set_callback(req, 0, NULL, NULL);
-
-               per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
-       }
-       /* before setting tcp_md5sig_pool_populated, we must commit all writes
-        * to memory. See smp_rmb() in tcp_get_md5sig_pool()
-        */
-       smp_wmb();
-       /* Paired with READ_ONCE() from tcp_alloc_md5sig_pool()
-        * and tcp_get_md5sig_pool().
-       */
-       WRITE_ONCE(tcp_md5sig_pool_populated, true);
-}
-
-bool tcp_alloc_md5sig_pool(void)
-{
-       /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
-       if (unlikely(!READ_ONCE(tcp_md5sig_pool_populated))) {
-               mutex_lock(&tcp_md5sig_mutex);
-
-               if (!tcp_md5sig_pool_populated)
-                       __tcp_alloc_md5sig_pool();
+       size_t scratch_size;
+       int ret;
 
-               mutex_unlock(&tcp_md5sig_mutex);
+       scratch_size = sizeof(union tcp_md5sum_block) + sizeof(struct tcphdr);
+       ret = tcp_sigpool_alloc_ahash("md5", scratch_size);
+       if (ret >= 0) {
+               /* As long as any md5 sigpool was allocated, the return
+                * id would stay the same. Re-write the id only for the case
+                * when previously all MD5 keys were deleted and this call
+                * allocates the first MD5 key, which may return a different
+                * sigpool id than was used previously.
+                */
+               WRITE_ONCE(tcp_md5_sigpool_id, ret); /* Avoids the compiler potentially being smart here */
+               return 0;
        }
-       /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
-       return READ_ONCE(tcp_md5sig_pool_populated);
+       return ret;
 }
-EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
-
 
-/**
- *     tcp_get_md5sig_pool - get md5sig_pool for this user
- *
- *     We use percpu structure, so if we succeed, we exit with preemption
- *     and BH disabled, to make sure another thread or softirq handling
- *     wont try to get same context.
- */
-struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
+void tcp_md5_release_sigpool(void)
 {
-       local_bh_disable();
-
-       /* Paired with WRITE_ONCE() from __tcp_alloc_md5sig_pool() */
-       if (READ_ONCE(tcp_md5sig_pool_populated)) {
-               /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
-               smp_rmb();
-               return this_cpu_ptr(&tcp_md5sig_pool);
-       }
-       local_bh_enable();
-       return NULL;
+       tcp_sigpool_release(READ_ONCE(tcp_md5_sigpool_id));
 }
-EXPORT_SYMBOL(tcp_get_md5sig_pool);
 
-int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
-                         const struct sk_buff *skb, unsigned int header_len)
+void tcp_md5_add_sigpool(void)
 {
-       struct scatterlist sg;
-       const struct tcphdr *tp = tcp_hdr(skb);
-       struct ahash_request *req = hp->md5_req;
-       unsigned int i;
-       const unsigned int head_data_len = skb_headlen(skb) > header_len ?
-                                          skb_headlen(skb) - header_len : 0;
-       const struct skb_shared_info *shi = skb_shinfo(skb);
-       struct sk_buff *frag_iter;
-
-       sg_init_table(&sg, 1);
-
-       sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
-       ahash_request_set_crypt(req, &sg, NULL, head_data_len);
-       if (crypto_ahash_update(req))
-               return 1;
-
-       for (i = 0; i < shi->nr_frags; ++i) {
-               const skb_frag_t *f = &shi->frags[i];
-               unsigned int offset = skb_frag_off(f);
-               struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
-
-               sg_set_page(&sg, page, skb_frag_size(f),
-                           offset_in_page(offset));
-               ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
-               if (crypto_ahash_update(req))
-                       return 1;
-       }
-
-       skb_walk_frags(skb, frag_iter)
-               if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
-                       return 1;
-
-       return 0;
+       tcp_sigpool_get(READ_ONCE(tcp_md5_sigpool_id));
 }
-EXPORT_SYMBOL(tcp_md5_hash_skb_data);
 
-int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
+int tcp_md5_hash_key(struct tcp_sigpool *hp,
+                    const struct tcp_md5sig_key *key)
 {
        u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
        struct scatterlist sg;
 
        sg_init_one(&sg, key->key, keylen);
-       ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
+       ahash_request_set_crypt(hp->req, &sg, NULL, keylen);
 
-       /* We use data_race() because tcp_md5_do_add() might change key->key under us */
-       return data_race(crypto_ahash_update(hp->md5_req));
+       /* We use data_race() because tcp_md5_do_add() might change
+        * key->key under us
+        */
+       return data_race(crypto_ahash_update(hp->req));
 }
 EXPORT_SYMBOL(tcp_md5_hash_key);
 
index 7583d4e34c8c5fedaf29177f458f7aa39ca0519a..7d81e90b6f5c4909687b424bde6784634b307c19 100644 (file)
@@ -1221,10 +1221,6 @@ static int __tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
        key = sock_kmalloc(sk, sizeof(*key), gfp | __GFP_ZERO);
        if (!key)
                return -ENOMEM;
-       if (!tcp_alloc_md5sig_pool()) {
-               sock_kfree_s(sk, key, sizeof(*key));
-               return -ENOMEM;
-       }
 
        memcpy(key->key, newkey, newkeylen);
        key->keylen = newkeylen;
@@ -1246,8 +1242,13 @@ int tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
        struct tcp_sock *tp = tcp_sk(sk);
 
        if (!rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk))) {
-               if (tcp_md5sig_info_add(sk, GFP_KERNEL))
+               if (tcp_md5_alloc_sigpool())
+                       return -ENOMEM;
+
+               if (tcp_md5sig_info_add(sk, GFP_KERNEL)) {
+                       tcp_md5_release_sigpool();
                        return -ENOMEM;
+               }
 
                if (!static_branch_inc(&tcp_md5_needed.key)) {
                        struct tcp_md5sig_info *md5sig;
@@ -1255,6 +1256,7 @@ int tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
                        md5sig = rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk));
                        rcu_assign_pointer(tp->md5sig_info, NULL);
                        kfree_rcu(md5sig, rcu);
+                       tcp_md5_release_sigpool();
                        return -EUSERS;
                }
        }
@@ -1271,8 +1273,12 @@ int tcp_md5_key_copy(struct sock *sk, const union tcp_md5_addr *addr,
        struct tcp_sock *tp = tcp_sk(sk);
 
        if (!rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk))) {
-               if (tcp_md5sig_info_add(sk, sk_gfp_mask(sk, GFP_ATOMIC)))
+               tcp_md5_add_sigpool();
+
+               if (tcp_md5sig_info_add(sk, sk_gfp_mask(sk, GFP_ATOMIC))) {
+                       tcp_md5_release_sigpool();
                        return -ENOMEM;
+               }
 
                if (!static_key_fast_inc_not_disabled(&tcp_md5_needed.key.key)) {
                        struct tcp_md5sig_info *md5sig;
@@ -1281,6 +1287,7 @@ int tcp_md5_key_copy(struct sock *sk, const union tcp_md5_addr *addr,
                        net_warn_ratelimited("Too many TCP-MD5 keys in the system\n");
                        rcu_assign_pointer(tp->md5sig_info, NULL);
                        kfree_rcu(md5sig, rcu);
+                       tcp_md5_release_sigpool();
                        return -EUSERS;
                }
        }
@@ -1380,7 +1387,7 @@ static int tcp_v4_parse_md5_keys(struct sock *sk, int optname,
                              cmd.tcpm_key, cmd.tcpm_keylen);
 }
 
-static int tcp_v4_md5_hash_headers(struct tcp_md5sig_pool *hp,
+static int tcp_v4_md5_hash_headers(struct tcp_sigpool *hp,
                                   __be32 daddr, __be32 saddr,
                                   const struct tcphdr *th, int nbytes)
 {
@@ -1400,38 +1407,35 @@ static int tcp_v4_md5_hash_headers(struct tcp_md5sig_pool *hp,
        _th->check = 0;
 
        sg_init_one(&sg, bp, sizeof(*bp) + sizeof(*th));
-       ahash_request_set_crypt(hp->md5_req, &sg, NULL,
+       ahash_request_set_crypt(hp->req, &sg, NULL,
                                sizeof(*bp) + sizeof(*th));
-       return crypto_ahash_update(hp->md5_req);
+       return crypto_ahash_update(hp->req);
 }
 
 static int tcp_v4_md5_hash_hdr(char *md5_hash, const struct tcp_md5sig_key *key,
                               __be32 daddr, __be32 saddr, const struct tcphdr *th)
 {
-       struct tcp_md5sig_pool *hp;
-       struct ahash_request *req;
+       struct tcp_sigpool hp;
 
-       hp = tcp_get_md5sig_pool();
-       if (!hp)
-               goto clear_hash_noput;
-       req = hp->md5_req;
+       if (tcp_sigpool_start(tcp_md5_sigpool_id, &hp))
+               goto clear_hash_nostart;
 
-       if (crypto_ahash_init(req))
+       if (crypto_ahash_init(hp.req))
                goto clear_hash;
-       if (tcp_v4_md5_hash_headers(hp, daddr, saddr, th, th->doff << 2))
+       if (tcp_v4_md5_hash_headers(&hp, daddr, saddr, th, th->doff << 2))
                goto clear_hash;
-       if (tcp_md5_hash_key(hp, key))
+       if (tcp_md5_hash_key(&hp, key))
                goto clear_hash;
-       ahash_request_set_crypt(req, NULL, md5_hash, 0);
-       if (crypto_ahash_final(req))
+       ahash_request_set_crypt(hp.req, NULL, md5_hash, 0);
+       if (crypto_ahash_final(hp.req))
                goto clear_hash;
 
-       tcp_put_md5sig_pool();
+       tcp_sigpool_end(&hp);
        return 0;
 
 clear_hash:
-       tcp_put_md5sig_pool();
-clear_hash_noput:
+       tcp_sigpool_end(&hp);
+clear_hash_nostart:
        memset(md5_hash, 0, 16);
        return 1;
 }
@@ -1440,9 +1444,8 @@ int tcp_v4_md5_hash_skb(char *md5_hash, const struct tcp_md5sig_key *key,
                        const struct sock *sk,
                        const struct sk_buff *skb)
 {
-       struct tcp_md5sig_pool *hp;
-       struct ahash_request *req;
        const struct tcphdr *th = tcp_hdr(skb);
+       struct tcp_sigpool hp;
        __be32 saddr, daddr;
 
        if (sk) { /* valid for establish/request sockets */
@@ -1454,30 +1457,28 @@ int tcp_v4_md5_hash_skb(char *md5_hash, const struct tcp_md5sig_key *key,
                daddr = iph->daddr;
        }
 
-       hp = tcp_get_md5sig_pool();
-       if (!hp)
-               goto clear_hash_noput;
-       req = hp->md5_req;
+       if (tcp_sigpool_start(tcp_md5_sigpool_id, &hp))
+               goto clear_hash_nostart;
 
-       if (crypto_ahash_init(req))
+       if (crypto_ahash_init(hp.req))
                goto clear_hash;
 
-       if (tcp_v4_md5_hash_headers(hp, daddr, saddr, th, skb->len))
+       if (tcp_v4_md5_hash_headers(&hp, daddr, saddr, th, skb->len))
                goto clear_hash;
-       if (tcp_md5_hash_skb_data(hp, skb, th->doff << 2))
+       if (tcp_sigpool_hash_skb_data(&hp, skb, th->doff << 2))
                goto clear_hash;
-       if (tcp_md5_hash_key(hp, key))
+       if (tcp_md5_hash_key(&hp, key))
                goto clear_hash;
-       ahash_request_set_crypt(req, NULL, md5_hash, 0);
-       if (crypto_ahash_final(req))
+       ahash_request_set_crypt(hp.req, NULL, md5_hash, 0);
+       if (crypto_ahash_final(hp.req))
                goto clear_hash;
 
-       tcp_put_md5sig_pool();
+       tcp_sigpool_end(&hp);
        return 0;
 
 clear_hash:
-       tcp_put_md5sig_pool();
-clear_hash_noput:
+       tcp_sigpool_end(&hp);
+clear_hash_nostart:
        memset(md5_hash, 0, 16);
        return 1;
 }
@@ -2296,6 +2297,18 @@ static int tcp_v4_init_sock(struct sock *sk)
        return 0;
 }
 
+#ifdef CONFIG_TCP_MD5SIG
+static void tcp_md5sig_info_free_rcu(struct rcu_head *head)
+{
+       struct tcp_md5sig_info *md5sig;
+
+       md5sig = container_of(head, struct tcp_md5sig_info, rcu);
+       kfree(md5sig);
+       static_branch_slow_dec_deferred(&tcp_md5_needed);
+       tcp_md5_release_sigpool();
+}
+#endif
+
 void tcp_v4_destroy_sock(struct sock *sk)
 {
        struct tcp_sock *tp = tcp_sk(sk);
@@ -2320,10 +2333,12 @@ void tcp_v4_destroy_sock(struct sock *sk)
 #ifdef CONFIG_TCP_MD5SIG
        /* Clean up the MD5 key list, if any */
        if (tp->md5sig_info) {
+               struct tcp_md5sig_info *md5sig;
+
+               md5sig = rcu_dereference_protected(tp->md5sig_info, 1);
                tcp_clear_md5_list(sk);
-               kfree_rcu(rcu_dereference_protected(tp->md5sig_info, 1), rcu);
-               tp->md5sig_info = NULL;
-               static_branch_slow_dec_deferred(&tcp_md5_needed);
+               call_rcu(&md5sig->rcu, tcp_md5sig_info_free_rcu);
+               rcu_assign_pointer(tp->md5sig_info, NULL);
        }
 #endif
 
index ace806c5bd0cdc2ec79eca5065e125418dd0c5b1..3dcb3fc36e64f166fa8f0eed6099c95b9840db10 100644 (file)
@@ -261,10 +261,9 @@ static void tcp_time_wait_init(struct sock *sk, struct tcp_timewait_sock *tcptw)
                tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
                if (!tcptw->tw_md5_key)
                        return;
-               if (!tcp_alloc_md5sig_pool())
-                       goto out_free;
                if (!static_key_fast_inc_not_disabled(&tcp_md5_needed.key.key))
                        goto out_free;
+               tcp_md5_add_sigpool();
        }
        return;
 out_free:
@@ -349,16 +348,26 @@ void tcp_time_wait(struct sock *sk, int state, int timeo)
 }
 EXPORT_SYMBOL(tcp_time_wait);
 
+#ifdef CONFIG_TCP_MD5SIG
+static void tcp_md5_twsk_free_rcu(struct rcu_head *head)
+{
+       struct tcp_md5sig_key *key;
+
+       key = container_of(head, struct tcp_md5sig_key, rcu);
+       kfree(key);
+       static_branch_slow_dec_deferred(&tcp_md5_needed);
+       tcp_md5_release_sigpool();
+}
+#endif
+
 void tcp_twsk_destructor(struct sock *sk)
 {
 #ifdef CONFIG_TCP_MD5SIG
        if (static_branch_unlikely(&tcp_md5_needed.key)) {
                struct tcp_timewait_sock *twsk = tcp_twsk(sk);
 
-               if (twsk->tw_md5_key) {
-                       kfree_rcu(twsk->tw_md5_key, rcu);
-                       static_branch_slow_dec_deferred(&tcp_md5_needed);
-               }
+               if (twsk->tw_md5_key)
+                       call_rcu(&twsk->tw_md5_key->rcu, tcp_md5_twsk_free_rcu);
        }
 #endif
 }
diff --git a/net/ipv4/tcp_sigpool.c b/net/ipv4/tcp_sigpool.c
new file mode 100644 (file)
index 0000000..65a8eaa
--- /dev/null
@@ -0,0 +1,358 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include <crypto/hash.h>
+#include <linux/cpu.h>
+#include <linux/kref.h>
+#include <linux/module.h>
+#include <linux/mutex.h>
+#include <linux/percpu.h>
+#include <linux/workqueue.h>
+#include <net/tcp.h>
+
+static size_t __scratch_size;
+static DEFINE_PER_CPU(void __rcu *, sigpool_scratch);
+
+struct sigpool_entry {
+       struct crypto_ahash     *hash;
+       const char              *alg;
+       struct kref             kref;
+       uint16_t                needs_key:1,
+                               reserved:15;
+};
+
+#define CPOOL_SIZE (PAGE_SIZE / sizeof(struct sigpool_entry))
+static struct sigpool_entry cpool[CPOOL_SIZE];
+static unsigned int cpool_populated;
+static DEFINE_MUTEX(cpool_mutex);
+
+/* Slow-path */
+struct scratches_to_free {
+       struct rcu_head rcu;
+       unsigned int cnt;
+       void *scratches[];
+};
+
+static void free_old_scratches(struct rcu_head *head)
+{
+       struct scratches_to_free *stf;
+
+       stf = container_of(head, struct scratches_to_free, rcu);
+       while (stf->cnt--)
+               kfree(stf->scratches[stf->cnt]);
+       kfree(stf);
+}
+
+/**
+ * sigpool_reserve_scratch - re-allocates scratch buffer, slow-path
+ * @size: request size for the scratch/temp buffer
+ */
+static int sigpool_reserve_scratch(size_t size)
+{
+       struct scratches_to_free *stf;
+       size_t stf_sz = struct_size(stf, scratches, num_possible_cpus());
+       int cpu, err = 0;
+
+       lockdep_assert_held(&cpool_mutex);
+       if (__scratch_size >= size)
+               return 0;
+
+       stf = kmalloc(stf_sz, GFP_KERNEL);
+       if (!stf)
+               return -ENOMEM;
+       stf->cnt = 0;
+
+       size = max(size, __scratch_size);
+       cpus_read_lock();
+       for_each_possible_cpu(cpu) {
+               void *scratch, *old_scratch;
+
+               scratch = kmalloc_node(size, GFP_KERNEL, cpu_to_node(cpu));
+               if (!scratch) {
+                       err = -ENOMEM;
+                       break;
+               }
+
+               old_scratch = rcu_replace_pointer(per_cpu(sigpool_scratch, cpu),
+                                       scratch, lockdep_is_held(&cpool_mutex));
+               if (!cpu_online(cpu) || !old_scratch) {
+                       kfree(old_scratch);
+                       continue;
+               }
+               stf->scratches[stf->cnt++] = old_scratch;
+       }
+       cpus_read_unlock();
+       if (!err)
+               __scratch_size = size;
+
+       call_rcu(&stf->rcu, free_old_scratches);
+       return err;
+}
+
+static void sigpool_scratch_free(void)
+{
+       int cpu;
+
+       for_each_possible_cpu(cpu)
+               kfree(rcu_replace_pointer(per_cpu(sigpool_scratch, cpu),
+                                         NULL, lockdep_is_held(&cpool_mutex)));
+       __scratch_size = 0;
+}
+
+static int __cpool_try_clone(struct crypto_ahash *hash)
+{
+       struct crypto_ahash *tmp;
+
+       tmp = crypto_clone_ahash(hash);
+       if (IS_ERR(tmp))
+               return PTR_ERR(tmp);
+
+       crypto_free_ahash(tmp);
+       return 0;
+}
+
+static int __cpool_alloc_ahash(struct sigpool_entry *e, const char *alg)
+{
+       struct crypto_ahash *cpu0_hash;
+       int ret;
+
+       e->alg = kstrdup(alg, GFP_KERNEL);
+       if (!e->alg)
+               return -ENOMEM;
+
+       cpu0_hash = crypto_alloc_ahash(alg, 0, CRYPTO_ALG_ASYNC);
+       if (IS_ERR(cpu0_hash)) {
+               ret = PTR_ERR(cpu0_hash);
+               goto out_free_alg;
+       }
+
+       e->needs_key = crypto_ahash_get_flags(cpu0_hash) & CRYPTO_TFM_NEED_KEY;
+
+       ret = __cpool_try_clone(cpu0_hash);
+       if (ret)
+               goto out_free_cpu0_hash;
+       e->hash = cpu0_hash;
+       kref_init(&e->kref);
+       return 0;
+
+out_free_cpu0_hash:
+       crypto_free_ahash(cpu0_hash);
+out_free_alg:
+       kfree(e->alg);
+       e->alg = NULL;
+       return ret;
+}
+
+/**
+ * tcp_sigpool_alloc_ahash - allocates pool for ahash requests
+ * @alg: name of async hash algorithm
+ * @scratch_size: reserve a tcp_sigpool::scratch buffer of this size
+ */
+int tcp_sigpool_alloc_ahash(const char *alg, size_t scratch_size)
+{
+       int i, ret;
+
+       /* slow-path */
+       mutex_lock(&cpool_mutex);
+       ret = sigpool_reserve_scratch(scratch_size);
+       if (ret)
+               goto out;
+       for (i = 0; i < cpool_populated; i++) {
+               if (!cpool[i].alg)
+                       continue;
+               if (strcmp(cpool[i].alg, alg))
+                       continue;
+
+               if (kref_read(&cpool[i].kref) > 0)
+                       kref_get(&cpool[i].kref);
+               else
+                       kref_init(&cpool[i].kref);
+               ret = i;
+               goto out;
+       }
+
+       for (i = 0; i < cpool_populated; i++) {
+               if (!cpool[i].alg)
+                       break;
+       }
+       if (i >= CPOOL_SIZE) {
+               ret = -ENOSPC;
+               goto out;
+       }
+
+       ret = __cpool_alloc_ahash(&cpool[i], alg);
+       if (!ret) {
+               ret = i;
+               if (i == cpool_populated)
+                       cpool_populated++;
+       }
+out:
+       mutex_unlock(&cpool_mutex);
+       return ret;
+}
+EXPORT_SYMBOL_GPL(tcp_sigpool_alloc_ahash);
+
+static void __cpool_free_entry(struct sigpool_entry *e)
+{
+       crypto_free_ahash(e->hash);
+       kfree(e->alg);
+       memset(e, 0, sizeof(*e));
+}
+
+static void cpool_cleanup_work_cb(struct work_struct *work)
+{
+       bool free_scratch = true;
+       unsigned int i;
+
+       mutex_lock(&cpool_mutex);
+       for (i = 0; i < cpool_populated; i++) {
+               if (kref_read(&cpool[i].kref) > 0) {
+                       free_scratch = false;
+                       continue;
+               }
+               if (!cpool[i].alg)
+                       continue;
+               __cpool_free_entry(&cpool[i]);
+       }
+       if (free_scratch)
+               sigpool_scratch_free();
+       mutex_unlock(&cpool_mutex);
+}
+
+static DECLARE_WORK(cpool_cleanup_work, cpool_cleanup_work_cb);
+static void cpool_schedule_cleanup(struct kref *kref)
+{
+       schedule_work(&cpool_cleanup_work);
+}
+
+/**
+ * tcp_sigpool_release - decreases number of users for a pool. If it was
+ * the last user of the pool, releases any memory that was consumed.
+ * @id: tcp_sigpool that was previously allocated by tcp_sigpool_alloc_ahash()
+ */
+void tcp_sigpool_release(unsigned int id)
+{
+       if (WARN_ON_ONCE(id > cpool_populated || !cpool[id].alg))
+               return;
+
+       /* slow-path */
+       kref_put(&cpool[id].kref, cpool_schedule_cleanup);
+}
+EXPORT_SYMBOL_GPL(tcp_sigpool_release);
+
+/**
+ * tcp_sigpool_get - increases number of users (refcounter) for a pool
+ * @id: tcp_sigpool that was previously allocated by tcp_sigpool_alloc_ahash()
+ */
+void tcp_sigpool_get(unsigned int id)
+{
+       if (WARN_ON_ONCE(id > cpool_populated || !cpool[id].alg))
+               return;
+       kref_get(&cpool[id].kref);
+}
+EXPORT_SYMBOL_GPL(tcp_sigpool_get);
+
+int tcp_sigpool_start(unsigned int id, struct tcp_sigpool *c) __cond_acquires(RCU_BH)
+{
+       struct crypto_ahash *hash;
+
+       rcu_read_lock_bh();
+       if (WARN_ON_ONCE(id > cpool_populated || !cpool[id].alg)) {
+               rcu_read_unlock_bh();
+               return -EINVAL;
+       }
+
+       hash = crypto_clone_ahash(cpool[id].hash);
+       if (IS_ERR(hash)) {
+               rcu_read_unlock_bh();
+               return PTR_ERR(hash);
+       }
+
+       c->req = ahash_request_alloc(hash, GFP_ATOMIC);
+       if (!c->req) {
+               crypto_free_ahash(hash);
+               rcu_read_unlock_bh();
+               return -ENOMEM;
+       }
+       ahash_request_set_callback(c->req, 0, NULL, NULL);
+
+       /* Pairs with tcp_sigpool_reserve_scratch(), scratch area is
+        * valid (allocated) until tcp_sigpool_end().
+        */
+       c->scratch = rcu_dereference_bh(*this_cpu_ptr(&sigpool_scratch));
+       return 0;
+}
+EXPORT_SYMBOL_GPL(tcp_sigpool_start);
+
+void tcp_sigpool_end(struct tcp_sigpool *c) __releases(RCU_BH)
+{
+       struct crypto_ahash *hash = crypto_ahash_reqtfm(c->req);
+
+       rcu_read_unlock_bh();
+       ahash_request_free(c->req);
+       crypto_free_ahash(hash);
+}
+EXPORT_SYMBOL_GPL(tcp_sigpool_end);
+
+/**
+ * tcp_sigpool_algo - return algorithm of tcp_sigpool
+ * @id: tcp_sigpool that was previously allocated by tcp_sigpool_alloc_ahash()
+ * @buf: buffer to return name of algorithm
+ * @buf_len: size of @buf
+ */
+size_t tcp_sigpool_algo(unsigned int id, char *buf, size_t buf_len)
+{
+       if (WARN_ON_ONCE(id > cpool_populated || !cpool[id].alg))
+               return -EINVAL;
+
+       return strscpy(buf, cpool[id].alg, buf_len);
+}
+EXPORT_SYMBOL_GPL(tcp_sigpool_algo);
+
+/**
+ * tcp_sigpool_hash_skb_data - hash data in skb with initialized tcp_sigpool
+ * @hp: tcp_sigpool pointer
+ * @skb: buffer to add sign for
+ * @header_len: TCP header length for this segment
+ */
+int tcp_sigpool_hash_skb_data(struct tcp_sigpool *hp,
+                             const struct sk_buff *skb,
+                             unsigned int header_len)
+{
+       const unsigned int head_data_len = skb_headlen(skb) > header_len ?
+                                          skb_headlen(skb) - header_len : 0;
+       const struct skb_shared_info *shi = skb_shinfo(skb);
+       const struct tcphdr *tp = tcp_hdr(skb);
+       struct ahash_request *req = hp->req;
+       struct sk_buff *frag_iter;
+       struct scatterlist sg;
+       unsigned int i;
+
+       sg_init_table(&sg, 1);
+
+       sg_set_buf(&sg, ((u8 *)tp) + header_len, head_data_len);
+       ahash_request_set_crypt(req, &sg, NULL, head_data_len);
+       if (crypto_ahash_update(req))
+               return 1;
+
+       for (i = 0; i < shi->nr_frags; ++i) {
+               const skb_frag_t *f = &shi->frags[i];
+               unsigned int offset = skb_frag_off(f);
+               struct page *page;
+
+               page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
+               sg_set_page(&sg, page, skb_frag_size(f), offset_in_page(offset));
+               ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
+               if (crypto_ahash_update(req))
+                       return 1;
+       }
+
+       skb_walk_frags(skb, frag_iter)
+               if (tcp_sigpool_hash_skb_data(hp, frag_iter, 0))
+                       return 1;
+
+       return 0;
+}
+EXPORT_SYMBOL(tcp_sigpool_hash_skb_data);
+
+MODULE_LICENSE("GPL");
+MODULE_DESCRIPTION("Per-CPU pool of crypto requests");
index dc27988512a6bce228b2de28022a3e5f3664ba73..ee53dad20a598c2cd439c1290c2cbbdc8dc88ff0 100644 (file)
@@ -671,7 +671,7 @@ static int tcp_v6_parse_md5_keys(struct sock *sk, int optname,
                              cmd.tcpm_key, cmd.tcpm_keylen);
 }
 
-static int tcp_v6_md5_hash_headers(struct tcp_md5sig_pool *hp,
+static int tcp_v6_md5_hash_headers(struct tcp_sigpool *hp,
                                   const struct in6_addr *daddr,
                                   const struct in6_addr *saddr,
                                   const struct tcphdr *th, int nbytes)
@@ -692,39 +692,36 @@ static int tcp_v6_md5_hash_headers(struct tcp_md5sig_pool *hp,
        _th->check = 0;
 
        sg_init_one(&sg, bp, sizeof(*bp) + sizeof(*th));
-       ahash_request_set_crypt(hp->md5_req, &sg, NULL,
+       ahash_request_set_crypt(hp->req, &sg, NULL,
                                sizeof(*bp) + sizeof(*th));
-       return crypto_ahash_update(hp->md5_req);
+       return crypto_ahash_update(hp->req);
 }
 
 static int tcp_v6_md5_hash_hdr(char *md5_hash, const struct tcp_md5sig_key *key,
                               const struct in6_addr *daddr, struct in6_addr *saddr,
                               const struct tcphdr *th)
 {
-       struct tcp_md5sig_pool *hp;
-       struct ahash_request *req;
+       struct tcp_sigpool hp;
 
-       hp = tcp_get_md5sig_pool();
-       if (!hp)
-               goto clear_hash_noput;
-       req = hp->md5_req;
+       if (tcp_sigpool_start(tcp_md5_sigpool_id, &hp))
+               goto clear_hash_nostart;
 
-       if (crypto_ahash_init(req))
+       if (crypto_ahash_init(hp.req))
                goto clear_hash;
-       if (tcp_v6_md5_hash_headers(hp, daddr, saddr, th, th->doff << 2))
+       if (tcp_v6_md5_hash_headers(&hp, daddr, saddr, th, th->doff << 2))
                goto clear_hash;
-       if (tcp_md5_hash_key(hp, key))
+       if (tcp_md5_hash_key(&hp, key))
                goto clear_hash;
-       ahash_request_set_crypt(req, NULL, md5_hash, 0);
-       if (crypto_ahash_final(req))
+       ahash_request_set_crypt(hp.req, NULL, md5_hash, 0);
+       if (crypto_ahash_final(hp.req))
                goto clear_hash;
 
-       tcp_put_md5sig_pool();
+       tcp_sigpool_end(&hp);
        return 0;
 
 clear_hash:
-       tcp_put_md5sig_pool();
-clear_hash_noput:
+       tcp_sigpool_end(&hp);
+clear_hash_nostart:
        memset(md5_hash, 0, 16);
        return 1;
 }
@@ -734,10 +731,9 @@ static int tcp_v6_md5_hash_skb(char *md5_hash,
                               const struct sock *sk,
                               const struct sk_buff *skb)
 {
-       const struct in6_addr *saddr, *daddr;
-       struct tcp_md5sig_pool *hp;
-       struct ahash_request *req;
        const struct tcphdr *th = tcp_hdr(skb);
+       const struct in6_addr *saddr, *daddr;
+       struct tcp_sigpool hp;
 
        if (sk) { /* valid for establish/request sockets */
                saddr = &sk->sk_v6_rcv_saddr;
@@ -748,30 +744,28 @@ static int tcp_v6_md5_hash_skb(char *md5_hash,
                daddr = &ip6h->daddr;
        }
 
-       hp = tcp_get_md5sig_pool();
-       if (!hp)
-               goto clear_hash_noput;
-       req = hp->md5_req;
+       if (tcp_sigpool_start(tcp_md5_sigpool_id, &hp))
+               goto clear_hash_nostart;
 
-       if (crypto_ahash_init(req))
+       if (crypto_ahash_init(hp.req))
                goto clear_hash;
 
-       if (tcp_v6_md5_hash_headers(hp, daddr, saddr, th, skb->len))
+       if (tcp_v6_md5_hash_headers(&hp, daddr, saddr, th, skb->len))
                goto clear_hash;
-       if (tcp_md5_hash_skb_data(hp, skb, th->doff << 2))
+       if (tcp_sigpool_hash_skb_data(&hp, skb, th->doff << 2))
                goto clear_hash;
-       if (tcp_md5_hash_key(hp, key))
+       if (tcp_md5_hash_key(&hp, key))
                goto clear_hash;
-       ahash_request_set_crypt(req, NULL, md5_hash, 0);
-       if (crypto_ahash_final(req))
+       ahash_request_set_crypt(hp.req, NULL, md5_hash, 0);
+       if (crypto_ahash_final(hp.req))
                goto clear_hash;
 
-       tcp_put_md5sig_pool();
+       tcp_sigpool_end(&hp);
        return 0;
 
 clear_hash:
-       tcp_put_md5sig_pool();
-clear_hash_noput:
+       tcp_sigpool_end(&hp);
+clear_hash_nostart:
        memset(md5_hash, 0, 16);
        return 1;
 }