|
| 1 | +// SPDX-License-Identifier: GPL-2.0 |
| 2 | +#include <linux/module.h> |
| 3 | +#include <linux/kernel.h> |
| 4 | +#include <linux/string.h> |
| 5 | +#include <linux/socket.h> |
| 6 | +#include <linux/net.h> |
| 7 | +#include <linux/fs.h> |
| 8 | +#include <net/af_unix.h> |
| 9 | +#include <net/scm.h> |
| 10 | +#include <linux/init.h> |
| 11 | + |
| 12 | +#include "scm.h" |
| 13 | + |
| 14 | +unsigned int unix_tot_inflight; |
| 15 | +EXPORT_SYMBOL(unix_tot_inflight); |
| 16 | + |
| 17 | +LIST_HEAD(gc_inflight_list); |
| 18 | +EXPORT_SYMBOL(gc_inflight_list); |
| 19 | + |
| 20 | +DEFINE_SPINLOCK(unix_gc_lock); |
| 21 | +EXPORT_SYMBOL(unix_gc_lock); |
| 22 | + |
| 23 | +struct sock *unix_get_socket(struct file *filp) |
| 24 | +{ |
| 25 | + struct sock *u_sock = NULL; |
| 26 | + struct inode *inode = file_inode(filp); |
| 27 | + |
| 28 | + /* Socket ? */ |
| 29 | + if (S_ISSOCK(inode->i_mode) && !(filp->f_mode & FMODE_PATH)) { |
| 30 | + struct socket *sock = SOCKET_I(inode); |
| 31 | + struct sock *s = sock->sk; |
| 32 | + |
| 33 | + /* PF_UNIX ? */ |
| 34 | + if (s && sock->ops && sock->ops->family == PF_UNIX) |
| 35 | + u_sock = s; |
| 36 | + } else { |
| 37 | + /* Could be an io_uring instance */ |
| 38 | + u_sock = io_uring_get_socket(filp); |
| 39 | + } |
| 40 | + return u_sock; |
| 41 | +} |
| 42 | +EXPORT_SYMBOL(unix_get_socket); |
| 43 | + |
| 44 | +/* Keep the number of times in flight count for the file |
| 45 | + * descriptor if it is for an AF_UNIX socket. |
| 46 | + */ |
| 47 | +void unix_inflight(struct user_struct *user, struct file *fp) |
| 48 | +{ |
| 49 | + struct sock *s = unix_get_socket(fp); |
| 50 | + |
| 51 | + spin_lock(&unix_gc_lock); |
| 52 | + |
| 53 | + if (s) { |
| 54 | + struct unix_sock *u = unix_sk(s); |
| 55 | + |
| 56 | + if (atomic_long_inc_return(&u->inflight) == 1) { |
| 57 | + BUG_ON(!list_empty(&u->link)); |
| 58 | + list_add_tail(&u->link, &gc_inflight_list); |
| 59 | + } else { |
| 60 | + BUG_ON(list_empty(&u->link)); |
| 61 | + } |
| 62 | + unix_tot_inflight++; |
| 63 | + } |
| 64 | + user->unix_inflight++; |
| 65 | + spin_unlock(&unix_gc_lock); |
| 66 | +} |
| 67 | + |
| 68 | +void unix_notinflight(struct user_struct *user, struct file *fp) |
| 69 | +{ |
| 70 | + struct sock *s = unix_get_socket(fp); |
| 71 | + |
| 72 | + spin_lock(&unix_gc_lock); |
| 73 | + |
| 74 | + if (s) { |
| 75 | + struct unix_sock *u = unix_sk(s); |
| 76 | + |
| 77 | + BUG_ON(!atomic_long_read(&u->inflight)); |
| 78 | + BUG_ON(list_empty(&u->link)); |
| 79 | + |
| 80 | + if (atomic_long_dec_and_test(&u->inflight)) |
| 81 | + list_del_init(&u->link); |
| 82 | + unix_tot_inflight--; |
| 83 | + } |
| 84 | + user->unix_inflight--; |
| 85 | + spin_unlock(&unix_gc_lock); |
| 86 | +} |
| 87 | + |
| 88 | +/* |
| 89 | + * The "user->unix_inflight" variable is protected by the garbage |
| 90 | + * collection lock, and we just read it locklessly here. If you go |
| 91 | + * over the limit, there might be a tiny race in actually noticing |
| 92 | + * it across threads. Tough. |
| 93 | + */ |
| 94 | +static inline bool too_many_unix_fds(struct task_struct *p) |
| 95 | +{ |
| 96 | + struct user_struct *user = current_user(); |
| 97 | + |
| 98 | + if (unlikely(user->unix_inflight > task_rlimit(p, RLIMIT_NOFILE))) |
| 99 | + return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN); |
| 100 | + return false; |
| 101 | +} |
| 102 | + |
| 103 | +int unix_attach_fds(struct scm_cookie *scm, struct sk_buff *skb) |
| 104 | +{ |
| 105 | + int i; |
| 106 | + |
| 107 | + if (too_many_unix_fds(current)) |
| 108 | + return -ETOOMANYREFS; |
| 109 | + |
| 110 | + /* |
| 111 | + * Need to duplicate file references for the sake of garbage |
| 112 | + * collection. Otherwise a socket in the fps might become a |
| 113 | + * candidate for GC while the skb is not yet queued. |
| 114 | + */ |
| 115 | + UNIXCB(skb).fp = scm_fp_dup(scm->fp); |
| 116 | + if (!UNIXCB(skb).fp) |
| 117 | + return -ENOMEM; |
| 118 | + |
| 119 | + for (i = scm->fp->count - 1; i >= 0; i--) |
| 120 | + unix_inflight(scm->fp->user, scm->fp->fp[i]); |
| 121 | + return 0; |
| 122 | +} |
| 123 | +EXPORT_SYMBOL(unix_attach_fds); |
| 124 | + |
| 125 | +void unix_detach_fds(struct scm_cookie *scm, struct sk_buff *skb) |
| 126 | +{ |
| 127 | + int i; |
| 128 | + |
| 129 | + scm->fp = UNIXCB(skb).fp; |
| 130 | + UNIXCB(skb).fp = NULL; |
| 131 | + |
| 132 | + for (i = scm->fp->count-1; i >= 0; i--) |
| 133 | + unix_notinflight(scm->fp->user, scm->fp->fp[i]); |
| 134 | +} |
| 135 | +EXPORT_SYMBOL(unix_detach_fds); |
| 136 | + |
| 137 | +void unix_destruct_scm(struct sk_buff *skb) |
| 138 | +{ |
| 139 | + struct scm_cookie scm; |
| 140 | + |
| 141 | + memset(&scm, 0, sizeof(scm)); |
| 142 | + scm.pid = UNIXCB(skb).pid; |
| 143 | + if (UNIXCB(skb).fp) |
| 144 | + unix_detach_fds(&scm, skb); |
| 145 | + |
| 146 | + /* Alas, it calls VFS */ |
| 147 | + /* So fscking what? fput() had been SMP-safe since the last Summer */ |
| 148 | + scm_destroy(&scm); |
| 149 | + sock_wfree(skb); |
| 150 | +} |
| 151 | +EXPORT_SYMBOL(unix_destruct_scm); |
0 commit comments