unix: Add ioctl(SIOCUNIXGRABFDS) to grab files of receive queue skbs
Subsystems (4)
v2 cover unix: Add ioctl(SIOCUNIXGRABFDS) to grab files of receive queue skbs
When a fd owning a counter of some critical resource, say, of a mount, it's impossible to umount that mount and disconnect related block device. That fd may be contained in some unix socket receive queue skb.
Despite we have an interface for detecting such the sockets queues (/proc/[PID]/fdinfo/[fd] shows non-zero scm_fds count if so) and it's possible to kill that process to release the counter, the problem is that there may be several processes, and it's not a good thing to kill each of them.
This patch adds a simple interface to grab files from receive queue, so the caller may analyze them, and even do that recursively, if grabbed file is unix socket itself. So, the described above problem may be solved by this ioctl() in pair with pidfd_getfd().
Note, that the existing recvmsg(,,MSG_PEEK) is not suitable for that purpose, since it modifies peek offset inside socket, and this results in a problem in case of examined process uses peek offset itself. Additional ptrace freezing of that task plus ioctl(SO_PEEK_OFF) won't help too, since that socket may relate to several tasks, and there is no reliable and non-racy way to detect that. Also, if the caller of such trick will die, the examined task will remain frozen forever. The new suggested ioctl(SIOCUNIXGRABFDS) does not have such problems.
The realization of ioctl(SIOCUNIXGRABFDS) is pretty simple. The only interesting thing is protocol with userspace. Firstly, we let userspace to know the number of all files in receive queue skbs. Then we receive fds one by one starting from requested offset. We return number of received fds if there is a successfully received fd, and this number may be less in case of error or desired fds number lack. Userspace may detect that situations by comparison of returned value and out.nr_all minus in.nr_skip. Looking over different variant this one looks the best for me (I considered returning error in case of error and there is a received fd. Also I considered returning number of received files as one more member in struct unix_ioc_grab_fds).
v2: Add "[PATCH 1/2] fs: Export __receive_fd()" to make receive_fd_user() be visible in [2/2] patch, when af_unix is module.
---
Kirill Tkhai (2):
fs: Export __receive_fd()
af_unix: Add ioctl(SIOCUNIXGRABFDS) to grab files of receive queue skbs
fs/file.c | 1 + include/uapi/linux/un.h | 12 ++++++++ net/unix/af_unix.c | 70 +++++++++++++++++++++++++++++++++++++++++++++++ 3 files changed, 83 insertions(+)
--
PATCH v2 1/2 fs: Export __receive_fd()
This is needed to make receive_fd_user() available in modules, and it will be used in next patch.
---
v2: New
fs/file.c | 1 + 1 file changed, 1 insertion(+)
fs/file.c +1 -0
@@ -1181,6 +1181,7 @@ int __receive_fd(struct file *file, int __user *ufd, unsigned int o_flags)__receive_sock (file);return new_fd; } +EXPORT_SYMBOL_GPL (__receive_fd);int receive_fd_replace (int new_fd,struct file *file,unsigned int o_flags) {
RE: [PATCH v2 1/2] fs: Export __receive_fd()
From: Kirill Tkhai
Quoted 15 lines
Sent: 15 August 2022 22:15
This is needed to make
receive_fd_user()available in modules, and it will be used in next patch.Signed-off-by Kirill Tkhai <[email protected]>---
v2: Newfs/file.c | 1 + 1 file changed, 1 insertion(+)fs/file.c +0
-0
fs/file.c int __receive_fd(struct file *file, int __user *ufd, unsigned int o_flags) · line 1184
Quoted 4 lines
fs/file.c +1
-0@@ -1181,6 +1181,7 @@ int __receive_fd(struct file *file, int __user *ufd, unsigned int o_flags)__receive_sock (file);return new_fd; } +EXPORT_SYMBOL_GPL (__receive_fd);
It doesn't seem right (to me) to be exporting a function with a __ prefix.
David
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
RE: [PATCH v2 1/2] fs: Export __receive_fd()
From: David Laight <[email protected]> Date: Tue, 16 Aug 2022 08:03:14 +0000
Quoted 11 lines
From: Kirill Tkhai
Quoted 10 lines
Sent: 15 August 2022 22:15
This is needed to make
receive_fd_user()available in modules, and it will be used in next patch.Signed-off-by Kirill Tkhai <[email protected]>---
v2: Newfs/file.c | 1 + 1 file changed, 1 insertion(+)
Quoted 12 lines
fs/file.c int __receive_fd(struct file *file, int __user *ufd, unsigned int o_flags) · line 1184
Quoted 9 lines
fs/file.c +1
-0@@ -1181,6 +1181,7 @@ int __receive_fd(struct file *file, int __user *ufd, unsigned int o_flags)__receive_sock (file);return new_fd; } +EXPORT_SYMBOL_GPL (__receive_fd);It doesn't seem right (to me) to be exporting a function with a __ prefix.
+1.
Now receive_fd() has inline and it's the problem.
Can we avoid this by moving receive_fd() in fs/file.c without inline and exporting it?
RE: [PATCH v2 1/2] fs: Export __receive_fd()
From: Kuniyuki Iwashima
Quoted 28 lines
Sent: 16 August 2022 18:15
From: David Laight <[email protected]> Date: Tue, 16 Aug 2022 08:03:14 +0000Quoted 11 lines
From: Kirill Tkhai
Quoted 10 lines
Sent: 15 August 2022 22:15
This is needed to make
receive_fd_user()available in modules, and it will be used in next patch.Signed-off-by Kirill Tkhai <[email protected]>---
v2: Newfs/file.c | 1 + 1 file changed, 1 insertion(+)Quoted 12 lines
fs/file.c int __receive_fd(struct file *file, int __user *ufd, unsigned int o_flags) · line 1184
Quoted 9 lines
fs/file.c +1
-0@@ -1181,6 +1181,7 @@ int __receive_fd(struct file *file, int __user *ufd, unsigned int o_flags)__receive_sock (file);return new_fd; } +EXPORT_SYMBOL_GPL (__receive_fd);It doesn't seem right (to me) to be exporting a function with a __ prefix.
Quoted 4 lines
+1.
Nowreceive_fd()has inline and it's the problem.
Can we avoid this by movingreceive_fd()infs/file.cwithout inline and exporting it?
It looks like it is receive_fd_user() that should be made a real function and then exported.__receive_fd() can then be static.
The extra function call will be noise - and the compiler may well either tail-call it or inline different copies of __receive_fd() into the two callers.
David
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
RE: [PATCH v2 1/2] fs: Export __receive_fd()
From: David Laight <[email protected]> Date: Tue, 16 Aug 2022 17:29:53 +0000
Quoted 29 lines
From: Kuniyuki Iwashima
Quoted 28 lines
Sent: 16 August 2022 18:15
From: David Laight <[email protected]> Date: Tue, 16 Aug 2022 08:03:14 +0000Quoted 23 lines
From: Kirill Tkhai
Quoted 15 lines
Sent: 15 August 2022 22:15
This is needed to make
receive_fd_user()available in modules, and it will be used in next patch.Signed-off-by Kirill Tkhai <[email protected]>---
v2: Newfs/file.c | 1 + 1 file changed, 1 insertion(+)fs/file.c +0
-0fs/file.c int __receive_fd(struct file *file, int __user *ufd, unsigned int o_flags) · line 1184
Quoted 4 lines
fs/file.c +1
-0@@ -1181,6 +1181,7 @@ int __receive_fd(struct file *file, int __user *ufd, unsigned int o_flags)__receive_sock (file);return new_fd; } +EXPORT_SYMBOL_GPL (__receive_fd);It doesn't seem right (to me) to be exporting a function with a __ prefix.
Quoted 7 lines
Quoted 4 lines
+1.
Nowreceive_fd()has inline and it's the problem.
Can we avoid this by movingreceive_fd()infs/file.cwithout inline and exporting it?It looks like it is
receive_fd_user()that should be made a real function and then exported.
Right, I did wrong copy-and-paste :p
Quoted 4 lines
__receive_fd()can then be static.
The extra function call will be noise - and the compiler may well either tail-call it or inline different copies of__receive_fd()into the two callers.
Re: [PATCH v2 1/2] fs: Export __receive_fd()
In reply to David LaightQuoted 11 lines · David Laight
From: Kirill Tkhai
Quoted 10 lines
Sent: 15 August 2022 22:15
This is needed to make
receive_fd_user()available in modules, and it will be used in next patch.Signed-off-by Kirill Tkhai <[email protected]>---
v2: Newfs/file.c | 1 + 1 file changed, 1 insertion(+)
Quoted 12 lines
fs/file.c int __receive_fd(struct file *file, int __user *ufd, unsigned int o_flags) · line 1184
Quoted 9 lines
fs/file.c +1
-0@@ -1181,6 +1181,7 @@ int __receive_fd(struct file *file, int __user *ufd, unsigned int o_flags)__receive_sock (file);return new_fd; } +EXPORT_SYMBOL_GPL (__receive_fd);It doesn't seem right (to me) to be exporting a function with a __ prefix.
I don't think so.
$git grep "EXPORT_SYMBOL(__\|EXPORT_SYMBOL_GPL(__" | wc -l 1649
PATCH v2 1/2 fs: Export __receive_fd()
In reply to Kirill TkhaiThis is needed to make receive_fd_user() available in modules, and it will be used in next patch.
---
v2: New
fs/file.c | 1 + 1 file changed, 1 insertion(+)
fs/file.c +1 -0
@@ -1181,6 +1181,7 @@ int __receive_fd(struct file *file, int __user *ufd, unsigned int o_flags)__receive_sock (file);return new_fd; } +EXPORT_SYMBOL_GPL (__receive_fd);int receive_fd_replace (int new_fd,struct file *file,unsigned int o_flags) {
PATCH v2 2/2 af_unix: Add ioctl(SIOCUNIXGRABFDS) to grab files of receive queue skbs
In reply to Kirill TkhaiWhen a fd owning a counter of some critical resource, say, of a mount, it's impossible to umount that mount and disconnect related block device. That fd may be contained in some unix socket receive queue skb.
Despite we have an interface for detecting such the sockets queues (/proc/[PID]/fdinfo/[fd] shows non-zero scm_fds count if so) and it's possible to kill that process to release the counter, the problem is that there may be several processes, and it's not a good thing to kill each of them.
This patch adds a simple interface to grab files from receive queue, so the caller may analyze them, and even do that recursively, if grabbed file is unix socket itself. So, the described above problem may be solved by this ioctl() in pair with pidfd_getfd().
Note, that the existing recvmsg(,,MSG_PEEK) is not suitable for that purpose, since it modifies peek offset inside socket, and this results in a problem in case of examined process uses peek offset itself. Additional ptrace freezing of that task plus ioctl(SO_PEEK_OFF) won't help too, since that socket may relate to several tasks, and there is no reliable and non-racy way to detect that. Also, if the caller of such trick will die, the examined task will remain frozen forever. The new suggested ioctl(SIOCUNIXGRABFDS) does not have such problems.
The realization of ioctl(SIOCUNIXGRABFDS) is pretty simple. The only interesting thing is protocol with userspace. Firstly, we let userspace to know the number of all files in receive queue skbs. Then we receive fds one by one starting from requested offset. We return number of received fds if there is a successfully received fd, and this number may be less in case of error or desired fds number lack. Userspace may detect that situations by comparison of returned value and out.nr_all minus in.nr_skip. Looking over different variant this one looks the best for me (I considered returning error in case of error and there is a received fd. Also I considered returning number of received files as one more member in struct unix_ioc_grab_fds).
--- include/uapi/linux/un.h | 12 ++++++++ net/unix/af_unix.c | 70 +++++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 82 insertions(+)
include/uapi/linux/un.h +12 -0
@@ -11,6 +11,18 @@ struct sockaddr_un { char sun_path[UNIX_PATH_MAX]; /* pathname */ }; +struct unix_ioc_grab_fds { +struct { +int nr_grab; +int nr_skip; +int *fds; + } in; +struct { +int nr_all; + } out; +}; +#define SIOCUNIXFILE (SIOCPROTOPRIVATE + 0)/* open a socket file with O_PATH */ +#define SIOCUNIXGRABFDS (SIOCPROTOPRIVATE + 1)/* grab files from recv queue */ #endif /* _LINUX_UN_H */
net/unix/af_unix.c +70 -0
82 lines not on this page · raw message
PATCH v2 2/2 af_unix: Add ioctl(SIOCUNIXGRABFDS) to grab files of receive queue skbs
From: Kirill Tkhai <[email protected]> Date: Tue, 16 Aug 2022 00:22:07 +0300
Quoted 24 lines
When a fd owning a counter of some critical resource, say, of a mount, it's impossible to umount that mount and disconnect related block device. That fd may be contained in some unix socket receive queue skb.
Despite we have an interface for detecting such the sockets queues (/proc/[PID]/fdinfo/[fd] shows non-zero
scm_fdscount if so) and it's possible to kill that process to release the counter, the problem is that there may be several processes, and it's not a good thing to kill each of them.This patch adds a simple interface to grab files from receive queue, so the caller may analyze them, and even do that recursively, if grabbed file is unix socket itself. So, the described above problem may be solved by this
ioctl()in pair withpidfd_getfd().Note, that the existing
recvmsg(,,MSG_PEEK)is not suitable for that purpose, since it modifies peek offset inside socket, and this results in a problem in case of examined process uses peek offset itself. Additional ptrace freezing of that task plusioctl(SO_PEEK_OFF)won't help too, since that socket may relate to several tasks, and there is no reliable and non-racy way to detect that. Also, if the caller of such trick will die, the examined task will remain frozen forever. The new suggested ioctl(SIOCUNIXGRABFDS) does not have such problems.
Quoted 3 lines
The realization of ioctl(SIOCUNIXGRABFDS) is pretty simple. The only interesting thing is protocol with userspace. Firstly, we let userspace to know the number of all files in receive queue skbs.
Before calling SIOCUNIXGRABFDS, we have to know nr_all by reading /proc/[PID]/fdinfo/[fd] to properly set nr_grab and nr_skip?
I'd use the same interface so that we can get it by passing 0 to todo.
Quoted 21 lines
Then we receive
fds one by one starting from requested offset. We return number of received fds if there is a successfully received fd, and this number may be less in case of error or desired fds number lack. Userspace may detect that situations by comparison of returned value and out.nr_all minus in.nr_skip. Looking over different variant this one looks the best for me (I considered returning error in case of error and there is a received fd. Also I considered returning number of received files as one more member in structunix_ioc_grab_fds).Signed-off-by Kirill Tkhai <[email protected]>--- include/uapi/linux/un.h | 12 ++++++++ net/unix/af_unix.c | 70 +++++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 82 insertions(+)include/uapi/linux/un.h +0
-0
include/uapi/linux/un.h struct sockaddr_un { · line 17
Quoted 7 lines
include/uapi/linux/un.h +4
-0@@ -11,6 +11,18 @@ struct sockaddr_un {char sun_path[UNIX_PATH_MAX ];/* pathname */ }; +struct unix_ioc_grab_fds { +struct { +int nr_grab; +int nr_skip;
We can save the first validation by using unsigned int.
include/uapi/linux/un.h struct sockaddr_un { · line 21
Quoted 4 lines
include/uapi/linux/un.h +4
-0+int *fds; + } in; +struct { +int nr_all;
unsigned int?
Quoted 20 lines
include/uapi/linux/un.h +4
-0+ } out; +}; +#define SIOCUNIXFILE (SIOCPROTOPRIVATE + 0)/* open a socket file with O_PATH */ +#define SIOCUNIXGRABFDS (SIOCPROTOPRIVATE + 1)/* grab files from recv queue */ #endif /* _LINUX_UN_H */ net/unix/af_unix.c +5
-0@@ -3079,6 +3079,73 @@ static int unix_open_file(struct sock *sk)return fd; } +static int unix_ioc_grab_fds (struct sock *sk,struct unix_ioc_grab_fds __user *uarg) +{ +int i, todo, skip, count, all, err, done =0 ; +struct unix_sock *u =unix_sk (sk); +struct unix_ioc_grab_fds arg;
net/unix/af_unix.c static int unix_open_file(struct sock *sk) · line 3096
Quoted 10 lines
net/unix/af_unix.c +10
-0+struct sk_buff *skb =NULL ; +struct scm_fp_list *fp; + +if (copy_from_user (&arg, uarg,sizeof (arg))) +return -EFAULT ; + + skip = arg.in.nr_skip; + todo = arg.in.nr_grab; + +if (skip <0 || todo <=0 )
If we accept 0 as a special value for todo:
if (skip < 0 || todo < 0)
And, unsigned int saves this.
net/unix/af_unix.c static int unix_open_file(struct sock *sk) · line 3107
Quoted 11 lines
net/unix/af_unix.c +11
-0+return -EINVAL ; +if (mutex_lock_interruptible (&u->iolock)) +return -EINTR ; + + all =atomic_read (&u->scm_stat.nr_fds); + err = -EFAULT ; +/* Set uarg->out.nr_all before the first file is received. */ +if (put_user (all, &uarg->out.nr_all)) + goto unlock; + err =0 ; +if (all <= skip)
If we accept 0 as a special value for todo:
if (all <= skip || !todo)
Quoted 26 lines
net/unix/af_unix.c +26
-0+ goto unlock; +if (all - skip < todo) + todo = all - skip; +while (todo) { +spin_lock (&sk->sk_receive_queue.lock); +if (!skb) + skb =skb_peek (&sk->sk_receive_queue); +else + skb =skb_peek_next (skb, &sk->sk_receive_queue); +spin_unlock (&sk->sk_receive_queue.lock); + +if (!skb) + goto unlock; + + fp =UNIXCB (skb).fp; + count = fp->count; +if (skip >= count) { + skip -= count; +continue ; + } + +for (i = skip; i < count && todo; i++) { + err =receive_fd_user (fp->fp[i], &arg.in.fds[done],0 ); +if (err <0 ) + goto unlock; + done++;
net/unix/af_unix.c static int unix_open_file(struct sock *sk) · line 3143
Quoted 10 lines
net/unix/af_unix.c +10
-0+ todo--; + } + skip =0 ; + } +unlock: +mutex_unlock (&u->iolock); + +/* Return number of fds (non-error) if there is a received file. */ +if (done) +return done;
return err;
You set 0 to err before the loop.
Quoted 18 lines
net/unix/af_unix.c +8
-0+if (err <0 ) +return err; +return 0 ; +} +static int unix_ioctl (struct socket *sock,unsigned int cmd,unsigned long arg) {struct sock *sk = sock->sk;@@ -3113,6 +3180,9 @@ static int unix_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg) }break ;#endif +case SIOCUNIXGRABFDS : + err =unix_ioc_grab_fds (sk, (struct unix_ioc_grab_fds __user *)arg); +break ;default : err = -ENOIOCTLCMD ;break ;
Re: [PATCH v2 2/2] af_unix: Add ioctl(SIOCUNIXGRABFDS) to grab files of receive queue skbs
Quoted 30 lines · Kuniyuki Iwashima
From: Kirill Tkhai <[email protected]> Date: Tue, 16 Aug 2022 00:22:07 +0300Quoted 24 lines
When a fd owning a counter of some critical resource, say, of a mount, it's impossible to umount that mount and disconnect related block device. That fd may be contained in some unix socket receive queue skb.
Despite we have an interface for detecting such the sockets queues (/proc/[PID]/fdinfo/[fd] shows non-zero
scm_fdscount if so) and it's possible to kill that process to release the counter, the problem is that there may be several processes, and it's not a good thing to kill each of them.This patch adds a simple interface to grab files from receive queue, so the caller may analyze them, and even do that recursively, if grabbed file is unix socket itself. So, the described above problem may be solved by this
ioctl()in pair withpidfd_getfd().Note, that the existing
recvmsg(,,MSG_PEEK)is not suitable for that purpose, since it modifies peek offset inside socket, and this results in a problem in case of examined process uses peek offset itself. Additional ptrace freezing of that task plusioctl(SO_PEEK_OFF)won't help too, since that socket may relate to several tasks, and there is no reliable and non-racy way to detect that. Also, if the caller of such trick will die, the examined task will remain frozen forever. The new suggested ioctl(SIOCUNIXGRABFDS) does not have such problems.Quoted 3 lines
The realization of ioctl(SIOCUNIXGRABFDS) is pretty simple. The only interesting thing is protocol with userspace. Firstly, we let userspace to know the number of all files in receive queue skbs.
Quoted 5 lines
Before calling SIOCUNIXGRABFDS, we have to know
nr_allby reading /proc/[PID]/fdinfo/[fd] to properly setnr_grabandnr_skip?I'd use the same interface so that we can get it by passing 0 to todo.
Yes, this make sense
Quoted 24 lines
Quoted 24 lines
Then we receive
fds one by one starting from requested offset. We return number of received fds if there is a successfully received fd, and this number may be less in case of error or desired fds number lack. Userspace may detect that situations by comparison of returned value and out.nr_all minus in.nr_skip. Looking over different variant this one looks the best for me (I considered returning error in case of error and there is a received fd. Also I considered returning number of received files as one more member in structunix_ioc_grab_fds).Signed-off-by Kirill Tkhai <[email protected]>--- include/uapi/linux/un.h | 12 ++++++++ net/unix/af_unix.c | 70 +++++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 82 insertions(+)include/uapi/linux/un.h +0
-0@@ -11,6 +11,18 @@ struct sockaddr_un {char sun_path[UNIX_PATH_MAX ];/* pathname */ };
Quoted 6 lines
include/uapi/linux/un.h struct sockaddr_un { · line 17
Quoted 4 lines
include/uapi/linux/un.h +4
-0+struct unix_ioc_grab_fds { +struct { +int nr_grab; +int nr_skip;We can save the first validation by using unsigned int.
This int was for uniformity for fd type. But since this is not a strict connection, we may use unsigned
Quoted 121 lines
include/uapi/linux/un.h struct sockaddr_un { · line 21
Quoted 4 lines
include/uapi/linux/un.h +4
-0+int *fds; + } in; +struct { +int nr_all;unsigned int?
Quoted 20 lines
include/uapi/linux/un.h +4
-0+ } out; +}; +#define SIOCUNIXFILE (SIOCPROTOPRIVATE + 0)/* open a socket file with O_PATH */ +#define SIOCUNIXGRABFDS (SIOCPROTOPRIVATE + 1)/* grab files from recv queue */ #endif /* _LINUX_UN_H */ net/unix/af_unix.c +5
-0@@ -3079,6 +3079,73 @@ static int unix_open_file(struct sock *sk)return fd; } +static int unix_ioc_grab_fds (struct sock *sk,struct unix_ioc_grab_fds __user *uarg) +{ +int i, todo, skip, count, all, err, done =0 ; +struct unix_sock *u =unix_sk (sk); +struct unix_ioc_grab_fds arg;net/unix/af_unix.c static int unix_open_file(struct sock *sk) · line 3096
Quoted 10 lines
net/unix/af_unix.c +10
-0+struct sk_buff *skb =NULL ; +struct scm_fp_list *fp; + +if (copy_from_user (&arg, uarg,sizeof (arg))) +return -EFAULT ; + + skip = arg.in.nr_skip; + todo = arg.in.nr_grab; + +if (skip <0 || todo <=0 )If we accept 0 as a special value for todo:
if (skip < 0 || todo < 0)And, unsigned int saves this.
net/unix/af_unix.c static int unix_open_file(struct sock *sk) · line 3107
Quoted 11 lines
net/unix/af_unix.c +11
-0+return -EINVAL ; +if (mutex_lock_interruptible (&u->iolock)) +return -EINTR ; + + all =atomic_read (&u->scm_stat.nr_fds); + err = -EFAULT ; +/* Set uarg->out.nr_all before the first file is received. */ +if (put_user (all, &uarg->out.nr_all)) + goto unlock; + err =0 ; +if (all <= skip)If we accept 0 as a special value for todo:
if (all <= skip || !todo)Quoted 26 lines
net/unix/af_unix.c +26
-0+ goto unlock; +if (all - skip < todo) + todo = all - skip; +while (todo) { +spin_lock (&sk->sk_receive_queue.lock); +if (!skb) + skb =skb_peek (&sk->sk_receive_queue); +else + skb =skb_peek_next (skb, &sk->sk_receive_queue); +spin_unlock (&sk->sk_receive_queue.lock); + +if (!skb) + goto unlock; + + fp =UNIXCB (skb).fp; + count = fp->count; +if (skip >= count) { + skip -= count; +continue ; + } + +for (i = skip; i < count && todo; i++) { + err =receive_fd_user (fp->fp[i], &arg.in.fds[done],0 ); +if (err <0 ) + goto unlock; + done++;net/unix/af_unix.c static int unix_open_file(struct sock *sk) · line 3143
Quoted 10 lines
net/unix/af_unix.c +10
-0+ todo--; + } + skip =0 ; + } +unlock: +mutex_unlock (&u->iolock); + +/* Return number of fds (non-error) if there is a received file. */ +if (done) +return done;return err;You set 0 to err before the loop.
Quoted 18 lines
net/unix/af_unix.c +8
-0+if (err <0 ) +return err; +return 0 ; +} +static int unix_ioctl (struct socket *sock,unsigned int cmd,unsigned long arg) {struct sock *sk = sock->sk;@@ -3113,6 +3180,9 @@ static int unix_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg) }break ;#endif +case SIOCUNIXGRABFDS : + err =unix_ioc_grab_fds (sk, (struct unix_ioc_grab_fds __user *)arg); +break ;default : err = -ENOIOCTLCMD ;break ;
PATCH v2 2/2 af_unix: Add ioctl(SIOCUNIXGRABFDS) to grab files of receive queue skbs
In reply to Kirill TkhaiWhen a fd owning a counter of some critical resource, say, of a mount, it's impossible to umount that mount and disconnect related block device. That fd may be contained in some unix socket receive queue skb.
Despite we have an interface for detecting such the sockets queues (/proc/[PID]/fdinfo/[fd] shows non-zero scm_fds count if so) and it's possible to kill that process to release the counter, the problem is that there may be several processes, and it's not a good thing to kill each of them.
This patch adds a simple interface to grab files from receive queue, so the caller may analyze them, and even do that recursively, if grabbed file is unix socket itself. So, the described above problem may be solved by this ioctl() in pair with pidfd_getfd().
Note, that the existing recvmsg(,,MSG_PEEK) is not suitable for that purpose, since it modifies peek offset inside socket, and this results in a problem in case of examined process uses peek offset itself. Additional ptrace freezing of that task plus ioctl(SO_PEEK_OFF) won't help too, since that socket may relate to several tasks, and there is no reliable and non-racy way to detect that. Also, if the caller of such trick will die, the examined task will remain frozen forever. The new suggested ioctl(SIOCUNIXGRABFDS) does not have such problems.
The realization of ioctl(SIOCUNIXGRABFDS) is pretty simple. The only interesting thing is protocol with userspace. Firstly, we let userspace to know the number of all files in receive queue skbs. Then we receive fds one by one starting from requested offset. We return number of received fds if there is a successfully received fd, and this number may be less in case of error or desired fds number lack. Userspace may detect that situations by comparison of returned value and out.nr_all minus in.nr_skip. Looking over different variant this one looks the best for me (I considered returning error in case of error and there is a received fd. Also I considered returning number of received files as one more member in struct unix_ioc_grab_fds).
--- include/uapi/linux/un.h | 12 ++++++++ net/unix/af_unix.c | 70 +++++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 82 insertions(+)
include/uapi/linux/un.h +12 -0
@@ -11,6 +11,18 @@ struct sockaddr_un { char sun_path[UNIX_PATH_MAX]; /* pathname */ }; +struct unix_ioc_grab_fds { +struct { +int nr_grab; +int nr_skip; +int *fds; + } in; +struct { +int nr_all; + } out; +}; +#define SIOCUNIXFILE (SIOCPROTOPRIVATE + 0)/* open a socket file with O_PATH */ +#define SIOCUNIXGRABFDS (SIOCPROTOPRIVATE + 1)/* grab files from recv queue */ #endif /* _LINUX_UN_H */
net/unix/af_unix.c +70 -0
82 lines not on this page · raw message
Al Viro
maintainerQuoted 24 lines · Kirill Tkhai
When a fd owning a counter of some critical resource, say, of a mount, it's impossible to umount that mount and disconnect related block device. That fd may be contained in some unix socket receive queue skb.
Despite we have an interface for detecting such the sockets queues (/proc/[PID]/fdinfo/[fd] shows non-zero
scm_fdscount if so) and it's possible to kill that process to release the counter, the problem is that there may be several processes, and it's not a good thing to kill each of them.This patch adds a simple interface to grab files from receive queue, so the caller may analyze them, and even do that recursively, if grabbed file is unix socket itself. So, the described above problem may be solved by this
ioctl()in pair withpidfd_getfd().Note, that the existing
recvmsg(,,MSG_PEEK)is not suitable for that purpose, since it modifies peek offset inside socket, and this results in a problem in case of examined process uses peek offset itself. Additional ptrace freezing of that task plusioctl(SO_PEEK_OFF)won't help too, since that socket may relate to several tasks, and there is no reliable and non-racy way to detect that. Also, if the caller of such trick will die, the examined task will remain frozen forever. The new suggested ioctl(SIOCUNIXGRABFDS) does not have such problems.
Quoted 11 lines
The realization of ioctl(SIOCUNIXGRABFDS) is pretty simple. The only interesting thing is protocol with userspace. Firstly, we let userspace to know the number of all files in receive queue skbs. Then we receive fds one by one starting from requested offset. We return number of received fds if there is a successfully received fd, and this number may be less in case of error or desired fds number lack. Userspace may detect that situations by comparison of returned value and out.nr_all minus in.nr_skip. Looking over different variant this one looks the best for me (I considered returning error in case of error and there is a received fd. Also I considered returning number of received files as one more member in struct
unix_ioc_grab_fds).
IMO it's a bad interface. Take a good look at the logics in scan_children(); TCP_LISTEN side is there for reason. And your magical ioctl won't help with that case.
Quoted 37 lines · Al Viro
Quoted 24 lines · Kirill Tkhai
When a fd owning a counter of some critical resource, say, of a mount, it's impossible to umount that mount and disconnect related block device. That fd may be contained in some unix socket receive queue skb.
Despite we have an interface for detecting such the sockets queues (/proc/[PID]/fdinfo/[fd] shows non-zero
scm_fdscount if so) and it's possible to kill that process to release the counter, the problem is that there may be several processes, and it's not a good thing to kill each of them.This patch adds a simple interface to grab files from receive queue, so the caller may analyze them, and even do that recursively, if grabbed file is unix socket itself. So, the described above problem may be solved by this
ioctl()in pair withpidfd_getfd().Note, that the existing
recvmsg(,,MSG_PEEK)is not suitable for that purpose, since it modifies peek offset inside socket, and this results in a problem in case of examined process uses peek offset itself. Additional ptrace freezing of that task plusioctl(SO_PEEK_OFF)won't help too, since that socket may relate to several tasks, and there is no reliable and non-racy way to detect that. Also, if the caller of such trick will die, the examined task will remain frozen forever. The new suggested ioctl(SIOCUNIXGRABFDS) does not have such problems.Quoted 11 lines
The realization of ioctl(SIOCUNIXGRABFDS) is pretty simple. The only interesting thing is protocol with userspace. Firstly, we let userspace to know the number of all files in receive queue skbs. Then we receive fds one by one starting from requested offset. We return number of received fds if there is a successfully received fd, and this number may be less in case of error or desired fds number lack. Userspace may detect that situations by comparison of returned value and out.nr_all minus in.nr_skip. Looking over different variant this one looks the best for me (I considered returning error in case of error and there is a received fd. Also I considered returning number of received files as one more member in struct
unix_ioc_grab_fds).
Quoted 3 lines
IMO it's a bad interface. Take a good look at the logics in
scan_children();TCP_LISTENside is there for reason. And your magical ioctl won't help with that case.
Good catch, thank you! I'll extend it to handle that case too. And your words bring to mind that fdinfo's scm_fds is also wrong for TCP_LISTEN sockets. I've sent separate patch with you CC'ed to fix that problem.
Kirill