Skip to contentSkip to conversation

syscall: introduce sendfd() syscall (v.2)

PATCH In review Alex Dubov 3 participants 6 messages Dec 3, 2014 – Dec 7, 2014 1 message not in the archive

Subsystems (3)
lore
Share & tools
lore

Alex Dubov

author
#0 lore raw

I would like to present my second attempt at file descriptor duplication over Posix.1b real-time signal transport. All the constructive points raised in the previous discussion are believed to be addressed.

To this end, I would like to address some concerns raised in the preceding discussion:

1. Claim: signals as a transport would not scale

Each task_struct allocated by kernel has its own signal queue, reliable, when Posix.1b signals are concerned. This queue essentially serves as per-task mail box, enabling complex applications to send signals from each thread to each thread directly, with very low overhead, and thus avoid any shared contention points outright (originating task's pid is passed along with the siginfo data, so source based dispatching is perfectly possible).

Also, signals can be trivially integrated with other communication mediums, as signalfd() syscall is perfectly compatible with epoll.

2. Claim: adding new functionality to the signal transport will create new attack/DoS vectors.

Nothing can be further from truth.

2.a. If task A has sufficient capabilities to send signals to task B, then task A is already in position to do anything it wants with task B, including killing it outright.

2.b. Flood attacks on signal queues are not dangerous to the system, as signal queues are relatively shallow and consume little memory even when full. Compare with infamous "recursive fd" attack against AF_UNIX fd transport , which plagues application development to this day (due to safeguards introduced to alleviate it).

2.c. Natural decoupling of signal transport from vfs internals prevents any sort of "recursive fd" attacks altogether (it is even safe to send the signalfd() fd through - this can be considered a convenient feature to replicate signal delivery masks around; of course, the receiving task will only receive its own signals through it, peeking on other task's signals will not be possible).

3. Suggestion: new file desriptors should not appear in destination processes out of the blue.

3.a. To receive the signal, process must make non-trivial preparations ( manipulate signal masks, etc), which would only happen if certain signals are expected.

3.b. In present implementation, file desriptor is only created at the destination when destination task explictly elects to receive the associated signal info with sigtimedwait/signalfd. In the absence of destination task cooperation, the only overhead on the kernel side will be a single pair of ref_count increment/decrement, that is, completely negligible.

3.c. Due to the nature of siginfo delivery, operations on file descriptor table are completely safe and indistinguishable from a normal dup() system call.

I would appreciate any additional constructive criticism, as it is in my interest as well to end up with safe and simple solution. However, I would prefer the criticism to target particular technical shortcomings, and not be derived from personal preferences, if possible.

author
#1 lore raw

PATCH 1/1 syscall: introduce sendfd() syscall (v.2)

Present patch introduces exceptionally easy to use, low latency and low overhead mechanism for transferring file descriptors between cooperating processes:

int sendfd(pid_t pid, int sig, int fd)

Given a target process pid, the sendfd() will queue a real-time signal for delivery to task referenced by pid. If signal can be delivered to destination tasks and it chooses to collect the associated signal info, a new file descriptor will be created on its behalf, pointing to file originally referred by fd (the value of newly created file descriptor will be communicated as integer payload within the siginfo data).

---
 arch/x86/syscalls/syscall_32.tbl   |  2 +
 arch/x86/syscalls/syscall_64.tbl   |  1 +
 include/asm-generic/siginfo.h      |  1 +
 include/linux/syscalls.h           |  1 +
 include/uapi/asm-generic/siginfo.h |  1 +
 init/Kconfig                       | 11 +++++
 kernel/signal.c                    | 89 ++++++++++++++++++++++++++++++++++++++
 kernel/sys_ni.c                    |  3 ++
 8 files changed, 109 insertions(+)
arch/x86/syscalls/syscall_32.tbl +2 -0
index 9fe1b5d..e2782bd 100644
@@ -364,3 +364,5 @@
 355	i386	getrandom		sys_getrandom
 356	i386	memfd_create		sys_memfd_create
 357	i386	bpf			sys_bpf
+358	i386	sendfd			sys_sendfd
+
arch/x86/syscalls/syscall_64.tbl +1 -0
index 281150b..4d6b55d 100644
@@ -328,6 +328,7 @@
 319	common	memfd_create		sys_memfd_create
 320	common	kexec_file_load		sys_kexec_file_load
 321	common	bpf			sys_bpf
+322	common	sendfd			sys_sendfd
 
 #
 # x32-specific system call numbers start at 512 to avoid cache impact
include/asm-generic/siginfo.h +1 -0
index 3d1a3af..c8af06f 100644
@@ -12,6 +12,7 @@
 #define __SI_RT		(5 << 16)
 #define __SI_MESGQ	(6 << 16)
 #define __SI_SYS	(7 << 16)
+#define __SI_FILEP	(8 << 16)
 #define __SI_CODE(T,N)	((T) | ((N) & 0xffff))
 
 struct siginfo;
include/linux/syscalls.h +1 -0
index bda9b81..1871b72f 100644
@@ -877,4 +877,5 @@ asmlinkage long sys_seccomp(unsigned int op, unsigned int flags,
 asmlinkage long sys_getrandom(char __user *buf, size_t count,
 			      unsigned int flags);
 asmlinkage long sys_bpf(int cmd, union bpf_attr *attr, unsigned int size);
+asmlinkage long sys_sendfd(pid_t pid, int sig, int fd);
 #endif
include/uapi/asm-generic/siginfo.h +1 -0
index ba5be7f..a92e38e 100644
@@ -148,6 +148,7 @@ typedef struct siginfo {
 #define __SI_RT		0
 #define __SI_MESGQ	0
 #define __SI_SYS	0
+#define __SI_FILEP	0
 #define __SI_CODE(T,N)	(N)
 #endif
 
init/Kconfig +11 -0
index 2081a4d..6a62a44 100644
@@ -1505,6 +1505,17 @@ config SIGNALFD
 
 	  If unsure, say Y.
 
+config SENDFD
+	bool "Enable sendfd() system call" if EXPERT
+	default y
+	help
+	  Enable the sendfd() system call that allows rapid duplication
+	  of file descriptor across process boundaries. The target process
+	  will receive a duplicate file descriptor delivered with one of
+	  Posix.1b real-time signals.
+
+	  If unsure, say Y.
+
 config TIMERFD
 	bool "Enable timerfd() system call" if EXPERT
 	select ANON_INODES
kernel/signal.c +89 -0

119 lines not on this page · raw message

kernel/sys_ni.c +3 -0
index 02aa418..353cddb 100644
@@ -224,3 +224,6 @@ cond_syscall(sys_seccomp);
 
 /* access BPF programs and maps */
 cond_syscall(sys_bpf);
+
+/* send file descriptor to another process */
+cond_syscall(sys_sendfd);
Signature · 1 line
1.8.3.2

A message that is not in the archive (sent off-list or not yet received): CAE2SPAa3FZz6MH41UB-CmruFw0kCTq=KS5MkNXMaiEidpSGkSg@mail.gmail.com

author
#3 lore raw
Quoted 4 lines · Bastien ROUCARIES

See senfd recvfd in gnulib. It wirk even under solaris

What's so special about a thin wrapper around domain sockets/named fifos (solaris style)?

#4 lore raw
In reply to Alex Dubov
Quoted 4 lines

2.a. If task A has sufficient capabilities to send signals to task B, then task A is already in position to do anything it wants with task B, including killing it outright.

Not entirely true.

- We have securirty models like SELinux
- We have namespaces and being able to send an fd between namespaces is not quite as flexible as you would make it

I suspect therefore it needs security hooks but otherwise looks more sane than the current AF_UNIX approach.

author
#5 lore raw
Quoted 5 lines · One Thousand Gnomes
Quoted 4 lines

2.a. If task A has sufficient capabilities to send signals to task B, then task A is already in position to do anything it wants with task B, including killing it outright.

Quoted 9 lines

Not entirely true.

- We have securirty models like SELinux
- We have namespaces and being able to send an fd between namespaces is not quite as flexible as you would make it

I suspect therefore it needs security hooks but otherwise looks more sane than the current AF_UNIX approach.

The best part about signal transport compared to anything in net/ is that it adheres to very straightforward and simple API contract. That is, you can tweak it here and there and still keep everything working.

1. adding an additional capability flag to selinux does not appear to be that complicated (it's got 4 capabilities related to signal handling already, fifth is not going to make much difference)

2. sending fds between namespaces may be prohibited outright; this would not be an unreasonable prohibition. A more flexible model may also be feasible, but I wonder if necessary.

#6 lore raw
In reply to One Thousand Gnomes
Quoted 5 lines · One Thousand Gnomes
Quoted 4 lines

2.a. If task A has sufficient capabilities to send signals to task B, then task A is already in position to do anything it wants with task B, including killing it outright.

Quoted 8 lines

Not entirely true.

- We have securirty models like SELinux
- We have namespaces and being able to send an fd between namespaces is not quite as flexible as you would make it

I suspect therefore it needs security hooks but otherwise looks more sane than the current AF_UNIX approach.

The right test for "can do anything" is "can_ptrace()"...
Pavel

Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html