From 97af84a6bba2ab2b9c704c08e67de3b5ea551bb2 Mon Sep 17 00:00:00 2001 From: Kuniyuki Iwashima Date: Tue, 23 Jan 2024 09:08:53 -0800 Subject: af_unix: Do not use atomic ops for unix_sk(sk)->inflight. When touching unix_sk(sk)->inflight, we are always under spin_lock(&unix_gc_lock). Let's convert unix_sk(sk)->inflight to the normal unsigned long. Signed-off-by: Kuniyuki Iwashima Reviewed-by: Simon Horman Link: https://lore.kernel.org/r/20240123170856.41348-3-kuniyu@amazon.com Signed-off-by: Jakub Kicinski --- net/unix/af_unix.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'net/unix/af_unix.c') diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c index ac1f2bc18fc9..1e9378036dcc 100644 --- a/net/unix/af_unix.c +++ b/net/unix/af_unix.c @@ -993,11 +993,11 @@ static struct sock *unix_create1(struct net *net, struct socket *sock, int kern, sk->sk_write_space = unix_write_space; sk->sk_max_ack_backlog = net->unx.sysctl_max_dgram_qlen; sk->sk_destruct = unix_sock_destructor; - u = unix_sk(sk); + u = unix_sk(sk); + u->inflight = 0; u->path.dentry = NULL; u->path.mnt = NULL; spin_lock_init(&u->lock); - atomic_long_set(&u->inflight, 0); INIT_LIST_HEAD(&u->link); mutex_init(&u->iolock); /* single task reading lock */ mutex_init(&u->bindlock); /* single task binding lock */ -- cgit From d9f21b3613337b55cc9d4a6ead484dca68475143 Mon Sep 17 00:00:00 2001 From: Kuniyuki Iwashima Date: Tue, 23 Jan 2024 09:08:56 -0800 Subject: af_unix: Try to run GC async. If more than 16000 inflight AF_UNIX sockets exist and the garbage collector is not running, unix_(dgram|stream)_sendmsg() call unix_gc(). Also, they wait for unix_gc() to complete. In unix_gc(), all inflight AF_UNIX sockets are traversed at least once, and more if they are the GC candidate. Thus, sendmsg() significantly slows down with too many inflight AF_UNIX sockets. However, if a process sends data with no AF_UNIX FD, the sendmsg() call does not need to wait for GC. After this change, only the process that meets the condition below will be blocked under such a situation. 1) cmsg contains AF_UNIX socket 2) more than 32 AF_UNIX sent by the same user are still inflight Note that even a sendmsg() call that does not meet the condition but has AF_UNIX FD will be blocked later in unix_scm_to_skb() by the spinlock, but we allow that as a bonus for sane users. The results below are the time spent in unix_dgram_sendmsg() sending 1 byte of data with no FD 4096 times on a host where 32K inflight AF_UNIX sockets exist. Without series: the sane sendmsg() needs to wait gc unreasonably. $ sudo /usr/share/bcc/tools/funclatency -p 11165 unix_dgram_sendmsg Tracing 1 functions for "unix_dgram_sendmsg"... Hit Ctrl-C to end. ^C nsecs : count distribution [...] 524288 -> 1048575 : 0 | | 1048576 -> 2097151 : 3881 |****************************************| 2097152 -> 4194303 : 214 |** | 4194304 -> 8388607 : 1 | | avg = 1825567 nsecs, total: 7477526027 nsecs, count: 4096 With series: the sane sendmsg() can finish much faster. $ sudo /usr/share/bcc/tools/funclatency -p 8702 unix_dgram_sendmsg Tracing 1 functions for "unix_dgram_sendmsg"... Hit Ctrl-C to end. ^C nsecs : count distribution [...] 128 -> 255 : 0 | | 256 -> 511 : 4092 |****************************************| 512 -> 1023 : 2 | | 1024 -> 2047 : 0 | | 2048 -> 4095 : 0 | | 4096 -> 8191 : 1 | | 8192 -> 16383 : 1 | | avg = 410 nsecs, total: 1680510 nsecs, count: 4096 Signed-off-by: Kuniyuki Iwashima Link: https://lore.kernel.org/r/20240123170856.41348-6-kuniyu@amazon.com Signed-off-by: Jakub Kicinski --- net/unix/af_unix.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) (limited to 'net/unix/af_unix.c') diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c index 1e9378036dcc..1720419d93d6 100644 --- a/net/unix/af_unix.c +++ b/net/unix/af_unix.c @@ -1923,11 +1923,12 @@ static int unix_dgram_sendmsg(struct socket *sock, struct msghdr *msg, long timeo; int err; - wait_for_unix_gc(); err = scm_send(sock, msg, &scm, false); if (err < 0) return err; + wait_for_unix_gc(scm.fp); + err = -EOPNOTSUPP; if (msg->msg_flags&MSG_OOB) goto out; @@ -2199,11 +2200,12 @@ static int unix_stream_sendmsg(struct socket *sock, struct msghdr *msg, bool fds_sent = false; int data_len; - wait_for_unix_gc(); err = scm_send(sock, msg, &scm, false); if (err < 0) return err; + wait_for_unix_gc(scm.fp); + err = -EOPNOTSUPP; if (msg->msg_flags & MSG_OOB) { #if IS_ENABLED(CONFIG_AF_UNIX_OOB) -- cgit From 99a7a5b9943ea2d05fb0dee38e4ae2290477ed83 Mon Sep 17 00:00:00 2001 From: Kuniyuki Iwashima Date: Mon, 29 Jan 2024 11:04:35 -0800 Subject: af_unix: Remove CONFIG_UNIX_SCM. Originally, the code related to garbage collection was all in garbage.c. Commit f4e65870e5ce ("net: split out functions related to registering inflight socket files") moved some functions to scm.c for io_uring and added CONFIG_UNIX_SCM just in case AF_UNIX was built as module. However, since commit 97154bcf4d1b ("af_unix: Kconfig: make CONFIG_UNIX bool"), AF_UNIX is no longer built separately. Also, io_uring does not support SCM_RIGHTS now. Let's move the functions back to garbage.c Signed-off-by: Kuniyuki Iwashima Acked-by: Jens Axboe Link: https://lore.kernel.org/r/20240129190435.57228-4-kuniyu@amazon.com Signed-off-by: Jakub Kicinski --- net/unix/af_unix.c | 63 ++++++++++++++++++++++++++++++++++++++++++++++++++++-- 1 file changed, 61 insertions(+), 2 deletions(-) (limited to 'net/unix/af_unix.c') diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c index 1720419d93d6..1cfbc586adb4 100644 --- a/net/unix/af_unix.c +++ b/net/unix/af_unix.c @@ -118,8 +118,6 @@ #include #include -#include "scm.h" - static atomic_long_t unix_nr_socks; static struct hlist_head bsd_socket_buckets[UNIX_HASH_SIZE / 2]; static spinlock_t bsd_socket_locks[UNIX_HASH_SIZE / 2]; @@ -1790,6 +1788,52 @@ out: return err; } +/* The "user->unix_inflight" variable is protected by the garbage + * collection lock, and we just read it locklessly here. If you go + * over the limit, there might be a tiny race in actually noticing + * it across threads. Tough. + */ +static inline bool too_many_unix_fds(struct task_struct *p) +{ + struct user_struct *user = current_user(); + + if (unlikely(READ_ONCE(user->unix_inflight) > task_rlimit(p, RLIMIT_NOFILE))) + return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN); + return false; +} + +static int unix_attach_fds(struct scm_cookie *scm, struct sk_buff *skb) +{ + int i; + + if (too_many_unix_fds(current)) + return -ETOOMANYREFS; + + /* Need to duplicate file references for the sake of garbage + * collection. Otherwise a socket in the fps might become a + * candidate for GC while the skb is not yet queued. + */ + UNIXCB(skb).fp = scm_fp_dup(scm->fp); + if (!UNIXCB(skb).fp) + return -ENOMEM; + + for (i = scm->fp->count - 1; i >= 0; i--) + unix_inflight(scm->fp->user, scm->fp->fp[i]); + + return 0; +} + +static void unix_detach_fds(struct scm_cookie *scm, struct sk_buff *skb) +{ + int i; + + scm->fp = UNIXCB(skb).fp; + UNIXCB(skb).fp = NULL; + + for (i = scm->fp->count - 1; i >= 0; i--) + unix_notinflight(scm->fp->user, scm->fp->fp[i]); +} + static void unix_peek_fds(struct scm_cookie *scm, struct sk_buff *skb) { scm->fp = scm_fp_dup(UNIXCB(skb).fp); @@ -1837,6 +1881,21 @@ static void unix_peek_fds(struct scm_cookie *scm, struct sk_buff *skb) spin_unlock(&unix_gc_lock); } +static void unix_destruct_scm(struct sk_buff *skb) +{ + struct scm_cookie scm; + + memset(&scm, 0, sizeof(scm)); + scm.pid = UNIXCB(skb).pid; + if (UNIXCB(skb).fp) + unix_detach_fds(&scm, skb); + + /* Alas, it calls VFS */ + /* So fscking what? fput() had been SMP-safe since the last Summer */ + scm_destroy(&scm); + sock_wfree(skb); +} + static int unix_scm_to_skb(struct scm_cookie *scm, struct sk_buff *skb, bool send_fds) { int err = 0; -- cgit From 1fbfdfaa590248c1d86407f578e40e5c65136330 Mon Sep 17 00:00:00 2001 From: Kuniyuki Iwashima Date: Mon, 25 Mar 2024 13:24:11 -0700 Subject: af_unix: Allocate struct unix_vertex for each inflight AF_UNIX fd. We will replace the garbage collection algorithm for AF_UNIX, where we will consider each inflight AF_UNIX socket as a vertex and its file descriptor as an edge in a directed graph. This patch introduces a new struct unix_vertex representing a vertex in the graph and adds its pointer to struct unix_sock. When we send a fd using the SCM_RIGHTS message, we allocate struct scm_fp_list to struct scm_cookie in scm_fp_copy(). Then, we bump each refcount of the inflight fds' struct file and save them in scm_fp_list.fp. After that, unix_attach_fds() inexplicably clones scm_fp_list of scm_cookie and sets it to skb. (We will remove this part after replacing GC.) Here, we add a new function call in unix_attach_fds() to preallocate struct unix_vertex per inflight AF_UNIX fd and link each vertex to skb's scm_fp_list.vertices. When sendmsg() succeeds later, if the socket of the inflight fd is still not inflight yet, we will set the preallocated vertex to struct unix_sock.vertex and link it to a global list unix_unvisited_vertices under spin_lock(&unix_gc_lock). If the socket is already inflight, we free the preallocated vertex. This is to avoid taking the lock unnecessarily when sendmsg() could fail later. In the following patch, we will similarly allocate another struct per edge, which will finally be linked to the inflight socket's unix_vertex.edges. And then, we will count the number of edges as unix_vertex.out_degree. Signed-off-by: Kuniyuki Iwashima Acked-by: Paolo Abeni Link: https://lore.kernel.org/r/20240325202425.60930-2-kuniyu@amazon.com Signed-off-by: Jakub Kicinski --- net/unix/af_unix.c | 6 ++++++ 1 file changed, 6 insertions(+) (limited to 'net/unix/af_unix.c') diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c index 5b41e2321209..a3b25d311560 100644 --- a/net/unix/af_unix.c +++ b/net/unix/af_unix.c @@ -980,6 +980,7 @@ static struct sock *unix_create1(struct net *net, struct socket *sock, int kern, sk->sk_destruct = unix_sock_destructor; u = unix_sk(sk); u->inflight = 0; + u->vertex = NULL; u->path.dentry = NULL; u->path.mnt = NULL; spin_lock_init(&u->lock); @@ -1805,6 +1806,9 @@ static int unix_attach_fds(struct scm_cookie *scm, struct sk_buff *skb) for (i = scm->fp->count - 1; i >= 0; i--) unix_inflight(scm->fp->user, scm->fp->fp[i]); + if (unix_prepare_fpl(UNIXCB(skb).fp)) + return -ENOMEM; + return 0; } @@ -1815,6 +1819,8 @@ static void unix_detach_fds(struct scm_cookie *scm, struct sk_buff *skb) scm->fp = UNIXCB(skb).fp; UNIXCB(skb).fp = NULL; + unix_destroy_fpl(scm->fp); + for (i = scm->fp->count - 1; i >= 0; i--) unix_notinflight(scm->fp->user, scm->fp->fp[i]); } -- cgit From 42f298c06b30bfe0a8cbee5d38644e618699e26e Mon Sep 17 00:00:00 2001 From: Kuniyuki Iwashima Date: Mon, 25 Mar 2024 13:24:13 -0700 Subject: af_unix: Link struct unix_edge when queuing skb. Just before queuing skb with inflight fds, we call scm_stat_add(), which is a good place to set up the preallocated struct unix_vertex and struct unix_edge in UNIXCB(skb).fp. Then, we call unix_add_edges() and construct the directed graph as follows: 1. Set the inflight socket's unix_sock to unix_edge.predecessor. 2. Set the receiver's unix_sock to unix_edge.successor. 3. Set the preallocated vertex to inflight socket's unix_sock.vertex. 4. Link inflight socket's unix_vertex.entry to unix_unvisited_vertices. 5. Link unix_edge.vertex_entry to the inflight socket's unix_vertex.edges. Let's say we pass the fd of AF_UNIX socket A to B and the fd of B to C. The graph looks like this: +-------------------------+ | unix_unvisited_vertices | <-------------------------. +-------------------------+ | + | | +--------------+ +--------------+ | +--------------+ | | unix_sock A | <---. .---> | unix_sock B | <-|-. .---> | unix_sock C | | +--------------+ | | +--------------+ | | | +--------------+ | .-+ | vertex | | | .-+ | vertex | | | | | vertex | | | +--------------+ | | | +--------------+ | | | +--------------+ | | | | | | | | | | +--------------+ | | | +--------------+ | | | | '-> | unix_vertex | | | '-> | unix_vertex | | | | | +--------------+ | | +--------------+ | | | `---> | entry | +---------> | entry | +-' | | |--------------| | | |--------------| | | | edges | <-. | | | edges | <-. | | +--------------+ | | | +--------------+ | | | | | | | | | .----------------------' | | .----------------------' | | | | | | | | | +--------------+ | | | +--------------+ | | | | unix_edge | | | | | unix_edge | | | | +--------------+ | | | +--------------+ | | `-> | vertex_entry | | | `-> | vertex_entry | | | |--------------| | | |--------------| | | | predecessor | +---' | | predecessor | +---' | |--------------| | |--------------| | | successor | +-----' | successor | +-----' +--------------+ +--------------+ Henceforth, we denote such a graph as A -> B (-> C). Now, we can express all inflight fd graphs that do not contain embryo sockets. We will support the particular case later. Signed-off-by: Kuniyuki Iwashima Acked-by: Paolo Abeni Link: https://lore.kernel.org/r/20240325202425.60930-4-kuniyu@amazon.com Signed-off-by: Jakub Kicinski --- net/unix/af_unix.c | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) (limited to 'net/unix/af_unix.c') diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c index a3b25d311560..24adbc4d5188 100644 --- a/net/unix/af_unix.c +++ b/net/unix/af_unix.c @@ -1943,8 +1943,10 @@ static void scm_stat_add(struct sock *sk, struct sk_buff *skb) struct scm_fp_list *fp = UNIXCB(skb).fp; struct unix_sock *u = unix_sk(sk); - if (unlikely(fp && fp->count)) + if (unlikely(fp && fp->count)) { atomic_add(fp->count, &u->scm_stat.nr_fds); + unix_add_edges(fp, u); + } } static void scm_stat_del(struct sock *sk, struct sk_buff *skb) @@ -1952,8 +1954,10 @@ static void scm_stat_del(struct sock *sk, struct sk_buff *skb) struct scm_fp_list *fp = UNIXCB(skb).fp; struct unix_sock *u = unix_sk(sk); - if (unlikely(fp && fp->count)) + if (unlikely(fp && fp->count)) { atomic_sub(fp->count, &u->scm_stat.nr_fds); + unix_del_edges(fp); + } } /* -- cgit From aed6ecef55d70de3762ce41c561b7f547dbaf107 Mon Sep 17 00:00:00 2001 From: Kuniyuki Iwashima Date: Mon, 25 Mar 2024 13:24:17 -0700 Subject: af_unix: Save listener for embryo socket. This is a prep patch for the following change, where we need to fetch the listening socket from the successor embryo socket during GC. We add a new field to struct unix_sock to save a pointer to a listening socket. We set it when connect() creates a new socket, and clear it when accept() is called. Signed-off-by: Kuniyuki Iwashima Acked-by: Paolo Abeni Link: https://lore.kernel.org/r/20240325202425.60930-8-kuniyu@amazon.com Signed-off-by: Jakub Kicinski --- net/unix/af_unix.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) (limited to 'net/unix/af_unix.c') diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c index 24adbc4d5188..af74e7ebc35a 100644 --- a/net/unix/af_unix.c +++ b/net/unix/af_unix.c @@ -979,6 +979,7 @@ static struct sock *unix_create1(struct net *net, struct socket *sock, int kern, sk->sk_max_ack_backlog = net->unx.sysctl_max_dgram_qlen; sk->sk_destruct = unix_sock_destructor; u = unix_sk(sk); + u->listener = NULL; u->inflight = 0; u->vertex = NULL; u->path.dentry = NULL; @@ -1598,6 +1599,7 @@ restart: newsk->sk_type = sk->sk_type; init_peercred(newsk); newu = unix_sk(newsk); + newu->listener = other; RCU_INIT_POINTER(newsk->sk_wq, &newu->peer_wq); otheru = unix_sk(other); @@ -1693,8 +1695,8 @@ static int unix_accept(struct socket *sock, struct socket *newsock, int flags, bool kern) { struct sock *sk = sock->sk; - struct sock *tsk; struct sk_buff *skb; + struct sock *tsk; int err; err = -EOPNOTSUPP; @@ -1719,6 +1721,7 @@ static int unix_accept(struct socket *sock, struct socket *newsock, int flags, } tsk = skb->sk; + unix_sk(tsk)->listener = NULL; skb_free_datagram(sk, skb); wake_up_interruptible(&unix_sk(sk)->peer_wait); -- cgit From dcf70df2048d27c5d186f013f101a4aefd63aa41 Mon Sep 17 00:00:00 2001 From: Kuniyuki Iwashima Date: Mon, 25 Mar 2024 13:24:18 -0700 Subject: af_unix: Fix up unix_edge.successor for embryo socket. To garbage collect inflight AF_UNIX sockets, we must define the cyclic reference appropriately. This is a bit tricky if the loop consists of embryo sockets. Suppose that the fd of AF_UNIX socket A is passed to D and the fd B to C and that C and D are embryo sockets of A and B, respectively. It may appear that there are two separate graphs, A (-> D) and B (-> C), but this is not correct. A --. .-- B X C <-' `-> D Now, D holds A's refcount, and C has B's refcount, so unix_release() will never be called for A and B when we close() them. However, no one can call close() for D and C to free skbs holding refcounts of A and B because C/D is in A/B's receive queue, which should have been purged by unix_release() for A and B. So, here's another type of cyclic reference. When a fd of an AF_UNIX socket is passed to an embryo socket, the reference is indirectly held by its parent listening socket. .-> A .-> B | `- sk_receive_queue | `- sk_receive_queue | `- skb | `- skb | `- sk == C | `- sk == D | `- sk_receive_queue | `- sk_receive_queue | `- skb +---------' `- skb +-. | | `---------------------------------------------------------' Technically, the graph must be denoted as A <-> B instead of A (-> D) and B (-> C) to find such a cyclic reference without touching each socket's receive queue. .-> A --. .-- B <-. | X | == A <-> B `-- C <-' `-> D --' We apply this fixup during GC by fetching the real successor by unix_edge_successor(). When we call accept(), we clear unix_sock.listener under unix_gc_lock not to confuse GC. Signed-off-by: Kuniyuki Iwashima Acked-by: Paolo Abeni Link: https://lore.kernel.org/r/20240325202425.60930-9-kuniyu@amazon.com Signed-off-by: Jakub Kicinski --- net/unix/af_unix.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'net/unix/af_unix.c') diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c index af74e7ebc35a..ae77e2dc0dae 100644 --- a/net/unix/af_unix.c +++ b/net/unix/af_unix.c @@ -1721,7 +1721,7 @@ static int unix_accept(struct socket *sock, struct socket *newsock, int flags, } tsk = skb->sk; - unix_sk(tsk)->listener = NULL; + unix_update_edges(unix_sk(tsk)); skb_free_datagram(sk, skb); wake_up_interruptible(&unix_sk(sk)->peer_wait); -- cgit From 4090fa373f0e763c43610853d2774b5979915959 Mon Sep 17 00:00:00 2001 From: Kuniyuki Iwashima Date: Mon, 25 Mar 2024 13:24:24 -0700 Subject: af_unix: Replace garbage collection algorithm. If we find a dead SCC during iteration, we call unix_collect_skb() to splice all skb in the SCC to the global sk_buff_head, hitlist. After iterating all SCC, we unlock unix_gc_lock and purge the queue. Signed-off-by: Kuniyuki Iwashima Acked-by: Paolo Abeni Link: https://lore.kernel.org/r/20240325202425.60930-15-kuniyu@amazon.com Signed-off-by: Jakub Kicinski --- net/unix/af_unix.c | 12 ------------ 1 file changed, 12 deletions(-) (limited to 'net/unix/af_unix.c') diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c index ae77e2dc0dae..27ca50ab1cd1 100644 --- a/net/unix/af_unix.c +++ b/net/unix/af_unix.c @@ -980,12 +980,10 @@ static struct sock *unix_create1(struct net *net, struct socket *sock, int kern, sk->sk_destruct = unix_sock_destructor; u = unix_sk(sk); u->listener = NULL; - u->inflight = 0; u->vertex = NULL; u->path.dentry = NULL; u->path.mnt = NULL; spin_lock_init(&u->lock); - INIT_LIST_HEAD(&u->link); mutex_init(&u->iolock); /* single task reading lock */ mutex_init(&u->bindlock); /* single task binding lock */ init_waitqueue_head(&u->peer_wait); @@ -1793,8 +1791,6 @@ static inline bool too_many_unix_fds(struct task_struct *p) static int unix_attach_fds(struct scm_cookie *scm, struct sk_buff *skb) { - int i; - if (too_many_unix_fds(current)) return -ETOOMANYREFS; @@ -1806,9 +1802,6 @@ static int unix_attach_fds(struct scm_cookie *scm, struct sk_buff *skb) if (!UNIXCB(skb).fp) return -ENOMEM; - for (i = scm->fp->count - 1; i >= 0; i--) - unix_inflight(scm->fp->user, scm->fp->fp[i]); - if (unix_prepare_fpl(UNIXCB(skb).fp)) return -ENOMEM; @@ -1817,15 +1810,10 @@ static int unix_attach_fds(struct scm_cookie *scm, struct sk_buff *skb) static void unix_detach_fds(struct scm_cookie *scm, struct sk_buff *skb) { - int i; - scm->fp = UNIXCB(skb).fp; UNIXCB(skb).fp = NULL; unix_destroy_fpl(scm->fp); - - for (i = scm->fp->count - 1; i >= 0; i--) - unix_notinflight(scm->fp->user, scm->fp->fp[i]); } static void unix_peek_fds(struct scm_cookie *scm, struct sk_buff *skb) -- cgit From 1abe267f173eae7ae76cf56232292e9641eb652f Mon Sep 17 00:00:00 2001 From: Eric Dumazet Date: Thu, 28 Mar 2024 14:40:32 +0000 Subject: net: add sk_wake_async_rcu() helper While looking at UDP receive performance, I saw sk_wake_async() was no longer inlined. This matters at least on AMD Zen1-4 platforms (see SRSO) This might be because rcu_read_lock() and rcu_read_unlock() are no longer nops in recent kernels ? Add sk_wake_async_rcu() variant, which must be called from contexts already holding rcu lock. As SOCK_FASYNC is deprecated in modern days, use unlikely() to give a hint to the compiler. sk_wake_async_rcu() is properly inlined from __udp_enqueue_schedule_skb() and sock_def_readable(). Signed-off-by: Eric Dumazet Link: https://lore.kernel.org/r/20240328144032.1864988-5-edumazet@google.com Signed-off-by: Jakub Kicinski --- net/unix/af_unix.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'net/unix/af_unix.c') diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c index 27ca50ab1cd1..533fb682c954 100644 --- a/net/unix/af_unix.c +++ b/net/unix/af_unix.c @@ -546,7 +546,7 @@ static void unix_write_space(struct sock *sk) if (skwq_has_sleeper(wq)) wake_up_interruptible_sync_poll(&wq->wait, EPOLLOUT | EPOLLWRNORM | EPOLLWRBAND); - sk_wake_async(sk, SOCK_WAKE_SPACE, POLL_OUT); + sk_wake_async_rcu(sk, SOCK_WAKE_SPACE, POLL_OUT); } rcu_read_unlock(); } -- cgit From 7c349ed090318b1c88a3e5dff3b24f732296edce Mon Sep 17 00:00:00 2001 From: Kuniyuki Iwashima Date: Mon, 1 Apr 2024 10:31:24 -0700 Subject: af_unix: Remove scm_fp_dup() in unix_attach_fds(). When we passed fds, we used to bump each file's refcount twice in scm_fp_copy() and scm_fp_dup() before linking the socket to gc_inflight_list. This is because we incremented the inflight count of the socket and linked it to the list in advance before passing skb to the destination socket. Otherwise, the inflight socket could have been garbage-collected in a small race window between linking the socket to the list and queuing skb: CPU 1 : sendmsg(X) w/ A's fd CPU 2 : close(A) ----- ----- /* Here A's refcount is 1, and inflight count is 0 */ bump A's refcount to 2 in scm_fp_copy() bump A's inflight count to 1 link A to gc_inflight_list decrement A's refcount to 1 /* A's refcount == inflight count, thus A could be GC candidate */ start GC mark A as candidate purge A's receive queue queue skb w/ A's fd to X /* A is queued, but all data has been lost */ After commit 4090fa373f0e ("af_unix: Replace garbage collection algorithm."), we increment the inflight count and link the socket to the global list only when queuing the skb. The race no longer exists, so let's not clone the fd nor bump the count in unix_attach_fds(). Signed-off-by: Kuniyuki Iwashima Link: https://lore.kernel.org/r/20240401173125.92184-2-kuniyu@amazon.com Signed-off-by: Jakub Kicinski --- net/unix/af_unix.c | 9 ++------- 1 file changed, 2 insertions(+), 7 deletions(-) (limited to 'net/unix/af_unix.c') diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c index 533fb682c954..78be8b520cef 100644 --- a/net/unix/af_unix.c +++ b/net/unix/af_unix.c @@ -1794,13 +1794,8 @@ static int unix_attach_fds(struct scm_cookie *scm, struct sk_buff *skb) if (too_many_unix_fds(current)) return -ETOOMANYREFS; - /* Need to duplicate file references for the sake of garbage - * collection. Otherwise a socket in the fps might become a - * candidate for GC while the skb is not yet queued. - */ - UNIXCB(skb).fp = scm_fp_dup(scm->fp); - if (!UNIXCB(skb).fp) - return -ENOMEM; + UNIXCB(skb).fp = scm->fp; + scm->fp = NULL; if (unix_prepare_fpl(UNIXCB(skb).fp)) return -ENOMEM; -- cgit From 118f457da9ed58a79e24b73c2ef0aa1987241f0e Mon Sep 17 00:00:00 2001 From: Kuniyuki Iwashima Date: Mon, 1 Apr 2024 10:31:25 -0700 Subject: af_unix: Remove lock dance in unix_peek_fds(). In the previous GC implementation, the shape of the inflight socket graph was not expected to change while GC was in progress. MSG_PEEK was tricky because it could install inflight fd silently and transform the graph. Let's say we peeked a fd, which was a listening socket, and accept()ed some embryo sockets from it. The garbage collection algorithm would have been confused because the set of sockets visited in scan_inflight() would change within the same GC invocation. That's why we placed spin_lock(&unix_gc_lock) and spin_unlock() in unix_peek_fds() with a fat comment. In the new GC implementation, we no longer garbage-collect the socket if it exists in another queue, that is, if it has a bridge to another SCC. Also, accept() will require the lock if it has edges. Thus, we need not do the complicated lock dance. Signed-off-by: Kuniyuki Iwashima Link: https://lore.kernel.org/r/20240401173125.92184-3-kuniyu@amazon.com Signed-off-by: Jakub Kicinski --- net/unix/af_unix.c | 42 ------------------------------------------ 1 file changed, 42 deletions(-) (limited to 'net/unix/af_unix.c') diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c index 78be8b520cef..61ecfa9c9c6b 100644 --- a/net/unix/af_unix.c +++ b/net/unix/af_unix.c @@ -1814,48 +1814,6 @@ static void unix_detach_fds(struct scm_cookie *scm, struct sk_buff *skb) static void unix_peek_fds(struct scm_cookie *scm, struct sk_buff *skb) { scm->fp = scm_fp_dup(UNIXCB(skb).fp); - - /* - * Garbage collection of unix sockets starts by selecting a set of - * candidate sockets which have reference only from being in flight - * (total_refs == inflight_refs). This condition is checked once during - * the candidate collection phase, and candidates are marked as such, so - * that non-candidates can later be ignored. While inflight_refs is - * protected by unix_gc_lock, total_refs (file count) is not, hence this - * is an instantaneous decision. - * - * Once a candidate, however, the socket must not be reinstalled into a - * file descriptor while the garbage collection is in progress. - * - * If the above conditions are met, then the directed graph of - * candidates (*) does not change while unix_gc_lock is held. - * - * Any operations that changes the file count through file descriptors - * (dup, close, sendmsg) does not change the graph since candidates are - * not installed in fds. - * - * Dequeing a candidate via recvmsg would install it into an fd, but - * that takes unix_gc_lock to decrement the inflight count, so it's - * serialized with garbage collection. - * - * MSG_PEEK is special in that it does not change the inflight count, - * yet does install the socket into an fd. The following lock/unlock - * pair is to ensure serialization with garbage collection. It must be - * done between incrementing the file count and installing the file into - * an fd. - * - * If garbage collection starts after the barrier provided by the - * lock/unlock, then it will see the elevated refcount and not mark this - * as a candidate. If a garbage collection is already in progress - * before the file count was incremented, then the lock/unlock pair will - * ensure that garbage collection is finished before progressing to - * installing the fd. - * - * (*) A -> B where B is on the queue of A or B is on the queue of C - * which is on the queue of listening socket A. - */ - spin_lock(&unix_gc_lock); - spin_unlock(&unix_gc_lock); } static void unix_destruct_scm(struct sk_buff *skb) -- cgit From b46f4eaa4f0ec38909fb0072eea3aeddb32f954e Mon Sep 17 00:00:00 2001 From: Kuniyuki Iwashima Date: Fri, 5 Apr 2024 15:10:57 -0700 Subject: af_unix: Clear stale u->oob_skb. syzkaller started to report deadlock of unix_gc_lock after commit 4090fa373f0e ("af_unix: Replace garbage collection algorithm."), but it just uncovers the bug that has been there since commit 314001f0bf92 ("af_unix: Add OOB support"). The repro basically does the following. from socket import * from array import array c1, c2 = socketpair(AF_UNIX, SOCK_STREAM) c1.sendmsg([b'a'], [(SOL_SOCKET, SCM_RIGHTS, array("i", [c2.fileno()]))], MSG_OOB) c2.recv(1) # blocked as no normal data in recv queue c2.close() # done async and unblock recv() c1.close() # done async and trigger GC A socket sends its file descriptor to itself as OOB data and tries to receive normal data, but finally recv() fails due to async close(). The problem here is wrong handling of OOB skb in manage_oob(). When recvmsg() is called without MSG_OOB, manage_oob() is called to check if the peeked skb is OOB skb. In such a case, manage_oob() pops it out of the receive queue but does not clear unix_sock(sk)->oob_skb. This is wrong in terms of uAPI. Let's say we send "hello" with MSG_OOB, and "world" without MSG_OOB. The 'o' is handled as OOB data. When recv() is called twice without MSG_OOB, the OOB data should be lost. >>> from socket import * >>> c1, c2 = socketpair(AF_UNIX, SOCK_STREAM, 0) >>> c1.send(b'hello', MSG_OOB) # 'o' is OOB data 5 >>> c1.send(b'world') 5 >>> c2.recv(5) # OOB data is not received b'hell' >>> c2.recv(5) # OOB date is skipped b'world' >>> c2.recv(5, MSG_OOB) # This should return an error b'o' In the same situation, TCP actually returns -EINVAL for the last recv(). Also, if we do not clear unix_sk(sk)->oob_skb, unix_poll() always set EPOLLPRI even though the data has passed through by previous recv(). To avoid these issues, we must clear unix_sk(sk)->oob_skb when dequeuing it from recv queue. The reason why the old GC did not trigger the deadlock is because the old GC relied on the receive queue to detect the loop. When it is triggered, the socket with OOB data is marked as GC candidate because file refcount == inflight count (1). However, after traversing all inflight sockets, the socket still has a positive inflight count (1), thus the socket is excluded from candidates. Then, the old GC lose the chance to garbage-collect the socket. With the old GC, the repro continues to create true garbage that will never be freed nor detected by kmemleak as it's linked to the global inflight list. That's why we couldn't even notice the issue. Fixes: 314001f0bf92 ("af_unix: Add OOB support") Reported-by: syzbot+7f7f201cc2668a8fd169@syzkaller.appspotmail.com Closes: https://syzkaller.appspot.com/bug?extid=7f7f201cc2668a8fd169 Signed-off-by: Kuniyuki Iwashima Reviewed-by: Eric Dumazet Link: https://lore.kernel.org/r/20240405221057.2406-1-kuniyu@amazon.com Signed-off-by: Jakub Kicinski --- net/unix/af_unix.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) (limited to 'net/unix/af_unix.c') diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c index 5b41e2321209..d032eb5fa6df 100644 --- a/net/unix/af_unix.c +++ b/net/unix/af_unix.c @@ -2665,7 +2665,9 @@ static struct sk_buff *manage_oob(struct sk_buff *skb, struct sock *sk, } } else if (!(flags & MSG_PEEK)) { skb_unlink(skb, &sk->sk_receive_queue); - consume_skb(skb); + WRITE_ONCE(u->oob_skb, NULL); + if (!WARN_ON_ONCE(skb_unref(skb))) + kfree_skb(skb); skb = skb_peek(&sk->sk_receive_queue); } } -- cgit From 283454c8a123072e5c386a5a2b5fc576aa455b6f Mon Sep 17 00:00:00 2001 From: Kuniyuki Iwashima Date: Wed, 10 Apr 2024 10:10:15 -0700 Subject: af_unix: Call manage_oob() for every skb in unix_stream_read_generic(). When we call recv() for AF_UNIX socket, we first peek one skb and calls manage_oob() to check if the skb is sent with MSG_OOB. However, when we fetch the next (and the following) skb, manage_oob() is not called now, leading a wrong behaviour. Let's say a socket send()s "hello" with MSG_OOB and the peer tries to recv() 5 bytes with MSG_PEEK. Here, we should get only "hell" without 'o', but actually not: >>> from socket import * >>> c1, c2 = socketpair(AF_UNIX, SOCK_STREAM) >>> c1.send(b'hello', MSG_OOB) 5 >>> c2.recv(5, MSG_PEEK) b'hello' The first skb fills 4 bytes, and the next skb is peeked but not properly checked by manage_oob(). Let's move up the again label to call manage_oob() for evry skb. With this patch: >>> from socket import * >>> c1, c2 = socketpair(AF_UNIX, SOCK_STREAM) >>> c1.send(b'hello', MSG_OOB) 5 >>> c2.recv(5, MSG_PEEK) b'hell' Fixes: 314001f0bf92 ("af_unix: Add OOB support") Signed-off-by: Kuniyuki Iwashima Link: https://lore.kernel.org/r/20240410171016.7621-2-kuniyu@amazon.com Signed-off-by: Jakub Kicinski --- net/unix/af_unix.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'net/unix/af_unix.c') diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c index d032eb5fa6df..f297320438bf 100644 --- a/net/unix/af_unix.c +++ b/net/unix/af_unix.c @@ -2741,6 +2741,7 @@ redo: last = skb = skb_peek(&sk->sk_receive_queue); last_len = last ? last->len : 0; +again: #if IS_ENABLED(CONFIG_AF_UNIX_OOB) if (skb) { skb = manage_oob(skb, sk, flags, copied); @@ -2752,7 +2753,6 @@ redo: } } #endif -again: if (skb == NULL) { if (copied >= target) goto unlock; -- cgit From 22dd70eb2c3d754862964377a75abafd3167346b Mon Sep 17 00:00:00 2001 From: Kuniyuki Iwashima Date: Wed, 10 Apr 2024 10:10:16 -0700 Subject: af_unix: Don't peek OOB data without MSG_OOB. Currently, we can read OOB data without MSG_OOB by using MSG_PEEK when OOB data is sitting on the front row, which is apparently wrong. >>> from socket import * >>> c1, c2 = socketpair(AF_UNIX, SOCK_STREAM) >>> c1.send(b'a', MSG_OOB) 1 >>> c2.recv(1, MSG_PEEK | MSG_DONTWAIT) b'a' If manage_oob() is called when no data has been copied, we only check if the socket enables SO_OOBINLINE or MSG_PEEK is not used. Otherwise, the skb is returned as is. However, here we should return NULL if MSG_PEEK is set and no data has been copied. Also, in such a case, we should not jump to the redo label because we will be caught in the loop and hog the CPU until normal data comes in. Then, we need to handle skb == NULL case with the if-clause below the manage_oob() block. With this patch: >>> from socket import * >>> c1, c2 = socketpair(AF_UNIX, SOCK_STREAM) >>> c1.send(b'a', MSG_OOB) 1 >>> c2.recv(1, MSG_PEEK | MSG_DONTWAIT) Traceback (most recent call last): File "", line 1, in BlockingIOError: [Errno 11] Resource temporarily unavailable Fixes: 314001f0bf92 ("af_unix: Add OOB support") Signed-off-by: Kuniyuki Iwashima Link: https://lore.kernel.org/r/20240410171016.7621-3-kuniyu@amazon.com Signed-off-by: Jakub Kicinski --- net/unix/af_unix.c | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) (limited to 'net/unix/af_unix.c') diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c index f297320438bf..9a6ad5974dff 100644 --- a/net/unix/af_unix.c +++ b/net/unix/af_unix.c @@ -2663,7 +2663,9 @@ static struct sk_buff *manage_oob(struct sk_buff *skb, struct sock *sk, WRITE_ONCE(u->oob_skb, NULL); consume_skb(skb); } - } else if (!(flags & MSG_PEEK)) { + } else if (flags & MSG_PEEK) { + skb = NULL; + } else { skb_unlink(skb, &sk->sk_receive_queue); WRITE_ONCE(u->oob_skb, NULL); if (!WARN_ON_ONCE(skb_unref(skb))) @@ -2745,11 +2747,9 @@ again: #if IS_ENABLED(CONFIG_AF_UNIX_OOB) if (skb) { skb = manage_oob(skb, sk, flags, copied); - if (!skb) { + if (!skb && copied) { unix_state_unlock(sk); - if (copied) - break; - goto redo; + break; } } #endif -- cgit From fd86344823b521149bb31d91eba900ba3525efa6 Mon Sep 17 00:00:00 2001 From: Kuniyuki Iwashima Date: Fri, 12 Apr 2024 19:19:28 -0700 Subject: af_unix: Try not to hold unix_gc_lock during accept(). Commit dcf70df2048d ("af_unix: Fix up unix_edge.successor for embryo socket.") added spin_lock(&unix_gc_lock) in accept() path, and it caused regression in a stress test as reported by kernel test robot. If the embryo socket is not part of the inflight graph, we need not hold the lock. To decide that in O(1) time and avoid the regression in the normal use case, 1. add a new stat unix_sk(sk)->scm_stat.nr_unix_fds 2. count the number of inflight AF_UNIX sockets in the receive queue under unix_state_lock() 3. move unix_update_edges() call under unix_state_lock() 4. avoid locking if nr_unix_fds is 0 in unix_update_edges() Reported-by: kernel test robot Closes: https://lore.kernel.org/oe-lkp/202404101427.92a08551-oliver.sang@intel.com Signed-off-by: Kuniyuki Iwashima Link: https://lore.kernel.org/r/20240413021928.20946-1-kuniyu@amazon.com Signed-off-by: Paolo Abeni --- net/unix/af_unix.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'net/unix/af_unix.c') diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c index 142d210b5b03..ed16d5f66df8 100644 --- a/net/unix/af_unix.c +++ b/net/unix/af_unix.c @@ -1719,12 +1719,12 @@ static int unix_accept(struct socket *sock, struct socket *newsock, int flags, } tsk = skb->sk; - unix_update_edges(unix_sk(tsk)); skb_free_datagram(sk, skb); wake_up_interruptible(&unix_sk(sk)->peer_wait); /* attach accepted sock to socket */ unix_state_lock(tsk); + unix_update_edges(unix_sk(tsk)); newsock->state = SS_CONNECTED; unix_sock_inherit_flags(sock, newsock); sock_graft(tsk, newsock); -- cgit From 540bf24fba16b88c1b3b9353927204b4f1074e25 Mon Sep 17 00:00:00 2001 From: Breno Leitao Date: Thu, 9 May 2024 01:14:46 -0700 Subject: af_unix: Fix data races in unix_release_sock/unix_stream_sendmsg A data-race condition has been identified in af_unix. In one data path, the write function unix_release_sock() atomically writes to sk->sk_shutdown using WRITE_ONCE. However, on the reader side, unix_stream_sendmsg() does not read it atomically. Consequently, this issue is causing the following KCSAN splat to occur: BUG: KCSAN: data-race in unix_release_sock / unix_stream_sendmsg write (marked) to 0xffff88867256ddbb of 1 bytes by task 7270 on cpu 28: unix_release_sock (net/unix/af_unix.c:640) unix_release (net/unix/af_unix.c:1050) sock_close (net/socket.c:659 net/socket.c:1421) __fput (fs/file_table.c:422) __fput_sync (fs/file_table.c:508) __se_sys_close (fs/open.c:1559 fs/open.c:1541) __x64_sys_close (fs/open.c:1541) x64_sys_call (arch/x86/entry/syscall_64.c:33) do_syscall_64 (arch/x86/entry/common.c:?) entry_SYSCALL_64_after_hwframe (arch/x86/entry/entry_64.S:130) read to 0xffff88867256ddbb of 1 bytes by task 989 on cpu 14: unix_stream_sendmsg (net/unix/af_unix.c:2273) __sock_sendmsg (net/socket.c:730 net/socket.c:745) ____sys_sendmsg (net/socket.c:2584) __sys_sendmmsg (net/socket.c:2638 net/socket.c:2724) __x64_sys_sendmmsg (net/socket.c:2753 net/socket.c:2750 net/socket.c:2750) x64_sys_call (arch/x86/entry/syscall_64.c:33) do_syscall_64 (arch/x86/entry/common.c:?) entry_SYSCALL_64_after_hwframe (arch/x86/entry/entry_64.S:130) value changed: 0x01 -> 0x03 The line numbers are related to commit dd5a440a31fa ("Linux 6.9-rc7"). Commit e1d09c2c2f57 ("af_unix: Fix data races around sk->sk_shutdown.") addressed a comparable issue in the past regarding sk->sk_shutdown. However, it overlooked resolving this particular data path. This patch only offending unix_stream_sendmsg() function, since the other reads seem to be protected by unix_state_lock() as discussed in Link: https://lore.kernel.org/all/20240508173324.53565-1-kuniyu@amazon.com/ Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2") Signed-off-by: Breno Leitao Reviewed-by: Kuniyuki Iwashima Link: https://lore.kernel.org/r/20240509081459.2807828-1-leitao@debian.org Signed-off-by: Jakub Kicinski --- net/unix/af_unix.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'net/unix/af_unix.c') diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c index 9a6ad5974dff..e94839d89b09 100644 --- a/net/unix/af_unix.c +++ b/net/unix/af_unix.c @@ -2270,7 +2270,7 @@ static int unix_stream_sendmsg(struct socket *sock, struct msghdr *msg, goto out_err; } - if (sk->sk_shutdown & SEND_SHUTDOWN) + if (READ_ONCE(sk->sk_shutdown) & SEND_SHUTDOWN) goto pipe_err; while (sent < len) { -- cgit From 92ef0fd55ac80dfc2e4654edfe5d1ddfa6e070fe Mon Sep 17 00:00:00 2001 From: Jens Axboe Date: Thu, 9 May 2024 09:20:08 -0600 Subject: net: change proto and proto_ops accept type Rather than pass in flags, error pointer, and whether this is a kernel invocation or not, add a struct proto_accept_arg struct as the argument. This then holds all of these arguments, and prepares accept for being able to pass back more information. No functional changes in this patch. Acked-by: Jakub Kicinski Signed-off-by: Jens Axboe --- net/unix/af_unix.c | 21 ++++++++++----------- 1 file changed, 10 insertions(+), 11 deletions(-) (limited to 'net/unix/af_unix.c') diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c index dc1651541723..26e3b5f1ee46 100644 --- a/net/unix/af_unix.c +++ b/net/unix/af_unix.c @@ -755,7 +755,7 @@ static int unix_bind(struct socket *, struct sockaddr *, int); static int unix_stream_connect(struct socket *, struct sockaddr *, int addr_len, int flags); static int unix_socketpair(struct socket *, struct socket *); -static int unix_accept(struct socket *, struct socket *, int, bool); +static int unix_accept(struct socket *, struct socket *, struct proto_accept_arg *arg); static int unix_getname(struct socket *, struct sockaddr *, int); static __poll_t unix_poll(struct file *, struct socket *, poll_table *); static __poll_t unix_dgram_poll(struct file *, struct socket *, @@ -1689,19 +1689,18 @@ static void unix_sock_inherit_flags(const struct socket *old, set_bit(SOCK_PASSSEC, &new->flags); } -static int unix_accept(struct socket *sock, struct socket *newsock, int flags, - bool kern) +static int unix_accept(struct socket *sock, struct socket *newsock, + struct proto_accept_arg *arg) { struct sock *sk = sock->sk; struct sk_buff *skb; struct sock *tsk; - int err; - err = -EOPNOTSUPP; + arg->err = -EOPNOTSUPP; if (sock->type != SOCK_STREAM && sock->type != SOCK_SEQPACKET) goto out; - err = -EINVAL; + arg->err = -EINVAL; if (sk->sk_state != TCP_LISTEN) goto out; @@ -1709,12 +1708,12 @@ static int unix_accept(struct socket *sock, struct socket *newsock, int flags, * so that no locks are necessary. */ - skb = skb_recv_datagram(sk, (flags & O_NONBLOCK) ? MSG_DONTWAIT : 0, - &err); + skb = skb_recv_datagram(sk, (arg->flags & O_NONBLOCK) ? MSG_DONTWAIT : 0, + &arg->err); if (!skb) { /* This means receive shutdown. */ - if (err == 0) - err = -EINVAL; + if (arg->err == 0) + arg->err = -EINVAL; goto out; } @@ -1732,7 +1731,7 @@ static int unix_accept(struct socket *sock, struct socket *newsock, int flags, return 0; out: - return err; + return arg->err; } -- cgit From 9841991a446c87f90f66f4b9fee6fe934c1336a2 Mon Sep 17 00:00:00 2001 From: Kuniyuki Iwashima Date: Thu, 16 May 2024 22:48:35 +0900 Subject: af_unix: Update unix_sk(sk)->oob_skb under sk_receive_queue lock. Billy Jheng Bing-Jhong reported a race between __unix_gc() and queue_oob(). __unix_gc() tries to garbage-collect close()d inflight sockets, and then if the socket has MSG_OOB in unix_sk(sk)->oob_skb, GC will drop the reference and set NULL to it locklessly. However, the peer socket still can send MSG_OOB message and queue_oob() can update unix_sk(sk)->oob_skb concurrently, leading NULL pointer dereference. [0] To fix the issue, let's update unix_sk(sk)->oob_skb under the sk_receive_queue's lock and take it everywhere we touch oob_skb. Note that we defer kfree_skb() in manage_oob() to silence lockdep false-positive (See [1]). [0]: BUG: kernel NULL pointer dereference, address: 0000000000000008 PF: supervisor write access in kernel mode PF: error_code(0x0002) - not-present page PGD 8000000009f5e067 P4D 8000000009f5e067 PUD 9f5d067 PMD 0 Oops: 0002 [#1] PREEMPT SMP PTI CPU: 3 PID: 50 Comm: kworker/3:1 Not tainted 6.9.0-rc5-00191-gd091e579b864 #110 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.16.0-0-gd239552ce722-prebuilt.qemu.org 04/01/2014 Workqueue: events delayed_fput RIP: 0010:skb_dequeue (./include/linux/skbuff.h:2386 ./include/linux/skbuff.h:2402 net/core/skbuff.c:3847) Code: 39 e3 74 3e 8b 43 10 48 89 ef 83 e8 01 89 43 10 49 8b 44 24 08 49 c7 44 24 08 00 00 00 00 49 8b 14 24 49 c7 04 24 00 00 00 00 <48> 89 42 08 48 89 10 e8 e7 c5 42 00 4c 89 e0 5b 5d 41 5c c3 cc cc RSP: 0018:ffffc900001bfd48 EFLAGS: 00000002 RAX: 0000000000000000 RBX: ffff8880088f5ae8 RCX: 00000000361289f9 RDX: 0000000000000000 RSI: 0000000000000206 RDI: ffff8880088f5b00 RBP: ffff8880088f5b00 R08: 0000000000080000 R09: 0000000000000001 R10: 0000000000000003 R11: 0000000000000001 R12: ffff8880056b6a00 R13: ffff8880088f5280 R14: 0000000000000001 R15: ffff8880088f5a80 FS: 0000000000000000(0000) GS:ffff88807dd80000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000000008 CR3: 0000000006314000 CR4: 00000000007506f0 PKRU: 55555554 Call Trace: unix_release_sock (net/unix/af_unix.c:654) unix_release (net/unix/af_unix.c:1050) __sock_release (net/socket.c:660) sock_close (net/socket.c:1423) __fput (fs/file_table.c:423) delayed_fput (fs/file_table.c:444 (discriminator 3)) process_one_work (kernel/workqueue.c:3259) worker_thread (kernel/workqueue.c:3329 kernel/workqueue.c:3416) kthread (kernel/kthread.c:388) ret_from_fork (arch/x86/kernel/process.c:153) ret_from_fork_asm (arch/x86/entry/entry_64.S:257) Modules linked in: CR2: 0000000000000008 Link: https://lore.kernel.org/netdev/a00d3993-c461-43f2-be6d-07259c98509a@rbox.co/ [1] Fixes: 1279f9d9dec2 ("af_unix: Call kfree_skb() for dead unix_(sk)->oob_skb in GC.") Reported-by: Billy Jheng Bing-Jhong Signed-off-by: Kuniyuki Iwashima Link: https://lore.kernel.org/r/20240516134835.8332-1-kuniyu@amazon.com Signed-off-by: Paolo Abeni --- net/unix/af_unix.c | 28 ++++++++++++++++++++++------ 1 file changed, 22 insertions(+), 6 deletions(-) (limited to 'net/unix/af_unix.c') diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c index fa906ec5e657..ca101690e740 100644 --- a/net/unix/af_unix.c +++ b/net/unix/af_unix.c @@ -2171,13 +2171,15 @@ static int queue_oob(struct socket *sock, struct msghdr *msg, struct sock *other maybe_add_creds(skb, sock, other); skb_get(skb); + scm_stat_add(other, skb); + + spin_lock(&other->sk_receive_queue.lock); if (ousk->oob_skb) consume_skb(ousk->oob_skb); - WRITE_ONCE(ousk->oob_skb, skb); + __skb_queue_tail(&other->sk_receive_queue, skb); + spin_unlock(&other->sk_receive_queue.lock); - scm_stat_add(other, skb); - skb_queue_tail(&other->sk_receive_queue, skb); sk_send_sigurg(other); unix_state_unlock(other); other->sk_data_ready(other); @@ -2568,8 +2570,10 @@ static int unix_stream_recv_urg(struct unix_stream_read_state *state) mutex_lock(&u->iolock); unix_state_lock(sk); + spin_lock(&sk->sk_receive_queue.lock); if (sock_flag(sk, SOCK_URGINLINE) || !u->oob_skb) { + spin_unlock(&sk->sk_receive_queue.lock); unix_state_unlock(sk); mutex_unlock(&u->iolock); return -EINVAL; @@ -2581,6 +2585,8 @@ static int unix_stream_recv_urg(struct unix_stream_read_state *state) WRITE_ONCE(u->oob_skb, NULL); else skb_get(oob_skb); + + spin_unlock(&sk->sk_receive_queue.lock); unix_state_unlock(sk); chunk = state->recv_actor(oob_skb, 0, chunk, state); @@ -2609,6 +2615,10 @@ static struct sk_buff *manage_oob(struct sk_buff *skb, struct sock *sk, consume_skb(skb); skb = NULL; } else { + struct sk_buff *unlinked_skb = NULL; + + spin_lock(&sk->sk_receive_queue.lock); + if (skb == u->oob_skb) { if (copied) { skb = NULL; @@ -2620,13 +2630,19 @@ static struct sk_buff *manage_oob(struct sk_buff *skb, struct sock *sk, } else if (flags & MSG_PEEK) { skb = NULL; } else { - skb_unlink(skb, &sk->sk_receive_queue); + __skb_unlink(skb, &sk->sk_receive_queue); WRITE_ONCE(u->oob_skb, NULL); - if (!WARN_ON_ONCE(skb_unref(skb))) - kfree_skb(skb); + unlinked_skb = skb; skb = skb_peek(&sk->sk_receive_queue); } } + + spin_unlock(&sk->sk_receive_queue.lock); + + if (unlinked_skb) { + WARN_ON_ONCE(skb_unref(unlinked_skb)); + kfree_skb(unlinked_skb); + } } return skb; } -- cgit