syscall: Setuid/Setgid doesn't apply to all threads on Linux

[gopherbot]

by ziutek@Lnet.pl:

What steps will reproduce the problem?

Compile attached test code. Run it as root like this:

# GOMAXPROCS=4 ./test 65534 65534

and note output:

gorutine 1: uid=0 euid=0 gid=0 egid=0
gorutine 2: uid=0 euid=0 gid=0 egid=0
gorutine 3: uid=0 euid=0 gid=0 egid=0
gorutine 4: uid=0 euid=0 gid=0 egid=0
gorutine 5: uid=0 euid=0 gid=0 egid=0
gorutine 6: uid=0 euid=0 gid=0 egid=0
gorutine 7: uid=0 euid=0 gid=0 egid=0
gorutine 8: uid=0 euid=0 gid=0 egid=0
gorutine 9: uid=0 euid=0 gid=0 egid=0
gorutine 0: uid=65534 euid=65534 gid=65534 egid=65534
gorutine 1: uid=0 euid=0 gid=0 egid=0
gorutine 2: uid=0 euid=0 gid=0 egid=0
gorutine 3: uid=0 euid=0 gid=0 egid=0
gorutine 4: uid=0 euid=0 gid=0 egid=0
gorutine 5: uid=0 euid=0 gid=0 egid=0
gorutine 6: uid=0 euid=0 gid=0 egid=0
gorutine 7: uid=0 euid=0 gid=0 egid=0
gorutine 8: uid=0 euid=0 gid=0 egid=0
gorutine 9: uid=0 euid=0 gid=0 egid=0
gorutine 0: uid=65534 euid=65534 gid=65534 egid=65534

Use ps -efL during test execution and note output:

UID        PID  PPID   LWP  C NLWP STIME TTY          TIME CMD
nobody   26088 25928 26088  0   10 11:56 pts/1    00:00:00 ./test 65534 65534
root     26088 25928 26089  0   10 11:56 pts/1    00:00:00 ./test 65534 65534
root     26088 25928 26090  0   10 11:56 pts/1    00:00:00 ./test 65534 65534
root     26088 25928 26091  0   10 11:56 pts/1    00:00:00 ./test 65534 65534
root     26088 25928 26092  0   10 11:56 pts/1    00:00:00 ./test 65534 65534
root     26088 25928 26093  0   10 11:56 pts/1    00:00:00 ./test 65534 65534
root     26088 25928 26094  0   10 11:56 pts/1    00:00:00 ./test 65534 65534
root     26088 25928 26095  0   10 11:56 pts/1    00:00:00 ./test 65534 65534
root     26088 25928 26096  0   10 11:56 pts/1    00:00:00 ./test 65534 65534
root     26088 25928 26097  0   10 11:56 pts/1    00:00:00 ./test 65534 65534

What is the expected output?

All threads must have the same UID/GID: (65534, nobody user in my system).

Which compiler are you using (5g, 6g, 8g, gccgo)?

I tested this with 6g and 8g.

Which operating system are you using?

Linux (Debian 6.0 SID on i386, Ubuntu 10.10 on amd64)

Which revision are you using?  (hg identify)

d8ba80011a98 release/release.2011-01-20

Please provide any additional information below.

http://groups.google.com/group/golang-nuts/browse_thread/thread/59597aafdd84a0e

Attachments:

1. test.go (1067 bytes)

[gopherbot]

Comment 1 by m@capitanio.org:

Well, you just must just put the gorutine creation after the suid/sguid call.
Isn't that what you would actually expect?
...
    en = syscall.Setuid(uid)
    if en != 0 {
        fmt.Println("Setuid error:", os.Errno(en))
        os.Exit(1)
    }

    for ii := 1; ii < 10; ii++ {
        go printIds(ii)
        time.Sleep(1e8)
    }
    printIds(0)
sudo -i
export GOMAXPROCS=2
/tmp/setuid 65534 65534
gorutine 1: uid=65534 euid=65534 gid=65534 egid=65534
gorutine 2: uid=65534 euid=65534 gid=65534 egid=65534
gorutine 3: uid=65534 euid=65534 gid=65534 egid=65534
gorutine 4: uid=65534 euid=65534 gid=65534 egid=65534
gorutine 5: uid=65534 euid=65534 gid=65534 egid=65534
gorutine 6: uid=65534 euid=65534 gid=65534 egid=65534
gorutine 7: uid=65534 euid=65534 gid=65534 egid=65534
gorutine 8: uid=65534 euid=65534 gid=65534 egid=65534
gorutine 9: uid=65534 euid=65534 gid=65534 egid=65534
gorutine 0: uid=65534 euid=65534 gid=65534 egid=65534
gorutine 1: uid=65534 euid=65534 gid=65534 egid=65534
gorutine 2: uid=65534 euid=65534 gid=65534 egid=65534
gorutine 3: uid=65534 euid=65534 gid=65534 egid=65534
gorutine 4: uid=65534 euid=65534 gid=65534 egid=65534
gorutine 5: uid=65534 euid=65534 gid=65534 egid=65534
gorutine 6: uid=65534 euid=65534 gid=65534 egid=65534
gorutine 7: uid=65534 euid=65534 gid=65534 egid=65534
gorutine 8: uid=65534 euid=65534 gid=65534 egid=65534

[alberts]

Comment 2:

By the time main starts executing it might be too late to be sure that you cover all the
goroutines.
A package could do:
func init() {
    go someHousekeepingFunc()
}
which would then have the wrong uid/gid.

[alberts]

Comment 3:

A more concrete example of this is perhaps:
http://golang.org/src/pkg/time/tick.go
To be safe, one would have to check the code of every package you import to be sure that
you don't inadvertently call a function that starts a goroutine before you get round to
calling Setuid.

[gopherbot]

Comment 4 by m@capitanio.org:

You could use runtime.Goroutines() to wait if you are the only one.
If it isn't possible to terminate all "someHousekeepingFunc()" e.g.
with Goexit(), I think setuid gains you to security not much ...

[gopherbot]

Comment 5 by ziutek@Lnet.pl:

Example code which shows how I discovered this problem:
package main
import (
    "os"
    "net"
    "http"
    "fmt"
    "syscall"
    "log"
)
type Handler string
func (hh *Handler) ServeHTTP(con http.ResponseWriter, req *http.Request) {
    fmt.Fprint(con, "Hello! I am %s. My UID/GID is %d/%d. Bye!\n", *hh,
        syscall.Getuid(), syscall.Getgid())
}
func main() {
    // To listen on port 80 we need root privileges
    ls, err := net.Listen("tcp", "127.0.0.1:80")
    if err != nil {
        log.Exitln("Can't listen:", err)
    }
    // We don't need root privileges any more
    if en := syscall.Setgid(65534); en != 0 {
        log.Exitln("Setgid error:", os.Errno(en))
    }
    if en := syscall.Setuid(65534); en != 0 {
        log.Exitln("Setuid error:", os.Errno(en))
    }
    // Run http service without root privileges
    handler := Handler("Test handler")
    if err = http.Serve(ls, &handler); err != nil {
        log.Exitln("Http server:", err)
    }
}
Compile it and run as root. Then use ps -efL to show threads:
UID        PID  PPID   LWP  C NLWP STIME TTY          TIME CMD
nobody   32401 32333 32401  0    2 18:48 pts/2    00:00:00 ./http
root     32401 32333 32402  0    2 18:48 pts/2    00:00:00 ./http
There is two threads. But I didn't create any gorutine explicitly. Use curl to send
request to the application:
$ curl 127.0.0.1
Hello! I am Test handler. My UID/GID is 65534/65534. Bye!
Now ps shows:
UID        PID  PPID   LWP  C NLWP STIME TTY          TIME CMD
nobody   32401 32333 32401  0    3 18:48 pts/2    00:00:00 ./http
root     32401 32333 32402  0    3 18:48 pts/2    00:00:00 ./http
nobody   32401 32333 32406  0    3 18:49 pts/2    00:00:00 ./http
Use siege stress test:
$ siege 127.0.0.1 -c25 -d0 -t 10s
** SIEGE 2.70
** Preparing 25 concurrent users for battle.
The server is now under siege...
Lifting the server siege..      done.                                                   
 Transactions:             12543 hits
Availability:             100.00 %
Elapsed time:               9.99 secs
Data transferred:           0.65 MB
Response time:              0.02 secs
Transaction rate:        1255.56 trans/sec
Throughput:             0.06 MB/sec
Concurrency:               24.79
Successful transactions:       12544
Failed transactions:               0
Longest transaction:            0.04
Shortest transaction:           0.01

Now ps shows:
UID        PID  PPID   LWP  C NLWP STIME TTY          TIME CMD
nobody   32401 32333 32401  0   11 18:48 pts/2    00:00:01 ./http
root     32401 32333 32402  0   11 18:48 pts/2    00:00:01 ./http
nobody   32401 32333 32406  0   11 18:49 pts/2    00:00:01 ./http
nobody   32401 32333 32553  0   11 18:51 pts/2    00:00:00 ./http
nobody   32401 32333 32554  2   11 18:51 pts/2    00:00:01 ./http
root     32401 32333 32555  3   11 18:51 pts/2    00:00:01 ./http
root     32401 32333 32556  3   11 18:51 pts/2    00:00:01 ./http
root     32401 32333 32557  0   11 18:51 pts/2    00:00:00 ./http
root     32401 32333 32558  3   11 18:51 pts/2    00:00:01 ./http
nobody   32401 32333 32559  3   11 18:51 pts/2    00:00:01 ./http
nobody   32401 32333 32560  3   11 18:51 pts/2    00:00:01 ./http
Use curl a few times:
$ curl 127.0.0.1
Hello! I am Test handler. My UID/GID is 0/0. Bye!
$ curl 127.0.0.1
Hello! I am Test handler. My UID/GID is 0/0. Bye!
$ curl 127.0.0.1
Hello! I am Test handler. My UID/GID is 0/0. Bye!
It's nice! I talking with root on your server!

[gopherbot]

Comment 6 by m@capitanio.org:

Oh, you are right. There is currently no official way to use a net socket without
spawning 2nd goroutine (EpollWait):
goroutine 2 [3]:
runtime.entersyscall+0x28 /data4/soft/go/go/src/pkg/runtime/proc.c:577
    runtime.entersyscall()
syscall.Syscall6+0x5 /data4/soft/go/go/src/pkg/syscall/asm_linux_amd64.s:40
    syscall.Syscall6()
syscall.EpollWait+0x8d /data4/soft/go/go/src/pkg/syscall/zsyscall_linux_amd64.go:188
    syscall.EpollWait(0x7fda00000006, 0x7fda62d566a0, 0x100000001, 0xffffffff, 0xc, ...)
net.*pollster·WaitFD+0xfe /data4/soft/go/go/src/pkg/net/fd_linux.go:116
    net.*pollster·WaitFD(0x7fda62d564d0, 0x0, 0x0, 0x0, 0x0, ...)
net.*pollServer·Run+0xa3 /data4/soft/go/go/src/pkg/net/fd.go:207
    net.*pollServer·Run(0x7fda62d20600, 0x0)
runtime.goexit /data4/soft/go/go/src/pkg/runtime/proc.c:149
    runtime.goexit()
but opening a file descriptor works:
f, err := os.Open("/etc/shadow", os.O_RDONLY | os.O_SYNC , 0666)
fmt.Println(runtime.Goroutines())
... syscall.Setuid(uid) ... syscall.Setgid(gid)
var buf [20]byte
_, err = f.Read(buf[:]);
fmt.Println(buf)
time.Sleep(100e9)
./run
open /etc/shadow: permission denied
sudo ./run
1
[114 111 111 ...
ps -efL|grep run
mc       14097 24745 14097  0    1 19:36 pts/4    00:00:00 /run 1001 1001

[robpike]

Comment 7:

Owner changed to r...@golang.org.

Status changed to Accepted.

[rsc]

Comment 8:

The syscall is doing what it advertises: it invokes
the Linux system call.  And the Linux system call only
affects the calling thread (!), confirmed by reading the
sources.
I was surprised to find that setuid works when called
from a C Linux pthreads (NPTL) program, though.
So I investigated further.
It turns out that glibc's setuid sends a signal to every
other thread to cause them to invoke the system call too.
http://goo.gl/8zf3C - called by setuid
http://goo.gl/CcXyX - signal handler
I suppose Go is going to need to do this at some point,
as part of implementing os.Setuid, os.Setgid, etc.
What a crock.
For now you can work around this by calling runtime.LockOSThread.
That locks the goroutine onto its current OS thread, so that
it only runs in that thread and that thread only runs that goroutine.
Then you can call Setuid Setgid etc and also ForkExec.

Status changed to LongTerm.

[gopherbot]

Comment 9 by ziutek@Lnet.pl:

If I add runtime.LockOSThread in my second example, just before Setgid, I get strange
behavior:
michal@md-lap:~$ curl 127.0.0.1
Hello! I am Test handler. My UID/GID is 65534/0. Bye!
michal@md-lap:~$ curl 127.0.0.1
Hello! I am Test handler. My UID/GID is 0/65534. Bye!
michal@md-lap:~$ curl 127.0.0.1
Hello! I am Test handler. My UID/GID is 0/0. Bye!
michal@md-lap:~$ curl 127.0.0.1
Hello! I am Test handler. My UID/GID is 65534/65534. Bye!
michal@md-lap:~$ curl 127.0.0.1
Hello! I am Test handler. My UID/GID is 65534/65534. Bye!
michal@md-lap:~$ curl 127.0.0.1
Hello! I am Test handler. My UID/GID is 65534/0. Bye!
michal@md-lap:~$ curl 127.0.0.1
Hello! I am Test handler. My UID/GID is 65534/65534. Bye!
michal@md-lap:~$ curl 127.0.0.1
Hello! I am Test handler. My UID/GID is 0/0. Bye!
michal@md-lap:~$ curl 127.0.0.1
Hello! I am Test handler. My UID/GID is 65534/65534. Bye!
michal@md-lap:~$ curl 127.0.0.1
Hello! I am Test handler. My UID/GID is 0/0. Bye!
michal@md-lap:~$ curl 127.0.0.1
Hello! I am Test handler. My UID/GID is 0/65534. Bye!

Attachments:

1. http.go (969 bytes)

[gopherbot]

Comment 10 by ziutek@Lnet.pl:

Sorry. I had not noticed it before, but this is also when there is no
runtime.LockOSThread call.
This is probably due to rescheduling between syscall.Getuid() and syscall.Getgid().

[rsc]

Comment 11:

Yes, even with LockOSThread it only applies to the current goroutine.
If you check from another goroutine, you may or may not see the change.
Sorry.
Another possible workaround for your specific case is to use setcap(8)
to give the binary the CAP_NET_BIND_SERVICE capability only.

[gopherbot]

Comment 12 by ziutek@Lnet.pl:

I probably found workaround for my web application:
package main
import (
    "os"
    "net"
    "http"
    "fmt"
    "syscall"
    "runtime"
    "log"
)
func lockUidGid(new_uid, new_gid int) {
    runtime.LockOSThread()
    uid := syscall.Getuid()
    gid := syscall.Getgid()
    if  uid == new_uid && gid == new_gid {
        return
    }
    if en := syscall.Setgid(new_uid); en != 0 {
        log.Exitln("Setgid error:", os.Errno(en))
    }
    if en := syscall.Setuid(new_gid); en != 0 {
        log.Exitln("Setuid error:", os.Errno(en))
    }
}
type Handler string
func (hh *Handler) ServeHTTP(con http.ResponseWriter, req *http.Request) {
    lockUidGid(65534, 65534)
    fmt.Fprintf(con, "Hello! I am %s. My UID/GID is %d/%d. Bye!\n", *hh,
        syscall.Getuid(), syscall.Getgid())
}
func main() {
    // To listen on port 80 we need root privileges
    ls, err := net.Listen("tcp", "127.0.0.1:80")
    if err != nil {
        log.Exitln("Can't listen:", err)
    }
    // We don't need root privileges any more
    lockUidGid(65534, 65534)
    // Run http service without root privileges
    handler := Handler("Test handler")
    if err = http.Serve(ls, &handler); err != nil {
        log.Exitln("Http server:", err)
    }
}
After running siege there is no threads with root privileges:
$ siege 127.0.0.1 -c25 -d0 -t10s
** SIEGE 2.69
** Preparing 25 concurrent users for battle.
The server is now under siege...
Lifting the server siege..      done.
Transactions:              28782 hits
Availability:             100.00 %
Elapsed time:               9.79 secs
Data transferred:           1.59 MB
Response time:              0.01 secs
Transaction rate:        2939.94 trans/sec
Throughput:             0.16 MB/sec
Concurrency:               24.86
Successful transactions:       28782
Failed transactions:               0
Longest transaction:            0.04
Shortest transaction:           0.00
$ ps -efL|egrep 'http|UID'|egrep -v egrep
UID        PID  PPID   LWP  C NLWP STIME TTY          TIME CMD
nobody    4109  2928  4109  1   10 23:00 pts/1    00:00:04 ./http
nobody    4109  2928  4110  0   10 23:00 pts/1    00:00:02 ./http
nobody    4109  2928  4112  0   10 23:00 pts/1    00:00:02 ./http
nobody    4109  2928  4153  0   10 23:01 pts/1    00:00:01 ./http
nobody    4109  2928  4154  0   10 23:01 pts/1    00:00:00 ./http
nobody    4109  2928  4155  0   10 23:01 pts/1    00:00:00 ./http
nobody    4109  2928  4156  0   10 23:01 pts/1    00:00:01 ./http
nobody    4109  2928  4157  0   10 23:01 pts/1    00:00:00 ./http
nobody    4109  2928  4158  0   10 23:01 pts/1    00:00:01 ./http
nobody    4109  2928  4159  0   10 23:01 pts/1    00:00:01 ./http
nobody    4109  2928  4109  1   10 23:00 pts/1    00:00:04 ./http
nobody    4109  2928  4110  0   10 23:00 pts/1    00:00:02 ./http
nobody    4109  2928  4112  0   10 23:00 pts/1    00:00:02 ./http
nobody    4109  2928  4153  0   10 23:01 pts/1    00:00:01 ./http
nobody    4109  2928  4154  0   10 23:01 pts/1    00:00:00 ./http
nobody    4109  2928  4155  0   10 23:01 pts/1    00:00:00 ./http
nobody    4109  2928  4156  0   10 23:01 pts/1    00:00:01 ./http
nobody    4109  2928  4157  0   10 23:01 pts/1    00:00:00 ./http
nobody    4109  2928  4158  0   10 23:01 pts/1    00:00:01 ./http
nobody    4109  2928  4159  0   10 23:01 pts/1    00:00:01 ./http
Thanks!

[gopherbot]

Comment 13 by ziutek@Lnet.pl:

I probably found workaround for my web application:
package main
import (
    "os"
    "net"
    "http"
    "fmt"
    "syscall"
    "runtime"
    "log"
)
func lockUidGid(new_uid, new_gid int) {
    runtime.LockOSThread()
    uid := syscall.Getuid()
    gid := syscall.Getgid()
    if  uid == new_uid && gid == new_gid {
        return
    }
    if en := syscall.Setgid(new_uid); en != 0 {
        log.Exitln("Setgid error:", os.Errno(en))
    }
    if en := syscall.Setuid(new_gid); en != 0 {
        log.Exitln("Setuid error:", os.Errno(en))
    }
}
type Handler string
func (hh *Handler) ServeHTTP(con http.ResponseWriter, req *http.Request) {
    lockUidGid(65534, 65534)
    fmt.Fprintf(con, "Hello! I am %s. My UID/GID is %d/%d. Bye!\n", *hh,
        syscall.Getuid(), syscall.Getgid())
}
func main() {
    // To listen on port 80 we need root privileges
    ls, err := net.Listen("tcp", "127.0.0.1:80")
    if err != nil {
        log.Exitln("Can't listen:", err)
    }
    // We don't need root privileges any more
    lockUidGid(65534, 65534)
    // Run http service without root privileges
    handler := Handler("Test handler")
    if err = http.Serve(ls, &handler); err != nil {
        log.Exitln("Http server:", err)
    }
}
After running siege there is no threads with root privileges:
$ siege 127.0.0.1 -c25 -d0 -t10s
** SIEGE 2.69
** Preparing 25 concurrent users for battle.
The server is now under siege...
Lifting the server siege..      done.
Transactions:                  28782 hits
Availability:                 100.00 %
Elapsed time:                   9.79 secs
Data transferred:               1.59 MB
Response time:                  0.01 secs
Transaction rate:            2939.94 trans/sec
Throughput:                     0.16 MB/sec
Concurrency:                   24.86
Successful transactions:       28782
Failed transactions:               0
Longest transaction:            0.04
Shortest transaction:           0.00
$ ps -efL|egrep 'http|UID'|egrep -v egrep
UID        PID  PPID   LWP  C NLWP STIME TTY          TIME CMD
nobody    4109  2928  4109  1   10 23:00 pts/1    00:00:04 ./http
nobody    4109  2928  4110  0   10 23:00 pts/1    00:00:02 ./http
nobody    4109  2928  4112  0   10 23:00 pts/1    00:00:02 ./http
nobody    4109  2928  4153  0   10 23:01 pts/1    00:00:01 ./http
nobody    4109  2928  4154  0   10 23:01 pts/1    00:00:00 ./http
nobody    4109  2928  4155  0   10 23:01 pts/1    00:00:00 ./http
nobody    4109  2928  4156  0   10 23:01 pts/1    00:00:01 ./http
nobody    4109  2928  4157  0   10 23:01 pts/1    00:00:00 ./http
nobody    4109  2928  4158  0   10 23:01 pts/1    00:00:01 ./http
nobody    4109  2928  4159  0   10 23:01 pts/1    00:00:01 ./http
Thanks!

[rsc]

Comment 14:

It's still not guaranteed that future goroutines won't have
the original 0/0 uid/gid.  Obviously if all tasks have been
switched then you're safe but there's no guarantee that
will switch all the tasks.  Using the network capability is
much safer if you are worried about this kind of thing.
Russ

[gopherbot]

Comment 15 by m@capitanio.org:

>For now you can work around this by calling runtime.LockOSThread.
runtime.GOMAXPROCS(1)
runtime.LockOSThread()
ls, _ := net.Listen("tcp", "localhost:42")
syscall.Setuid(uid)
http.Serve(ls, nil)
Was that what you meant? It's actually impossible ;)
There is always +1 thread spawned by Listen that keeps
running as root.
ps -efL| grep setuid
mc       19213 24745 19213  0    2 22:41 pts/4    00:00:00 /tmp/setuid 1001 1001
root     19213 24745 19215  0    2 22:41 pts/4    00:00:00 /tmp/setuid 1001 1001
^\SIGQUIT: quit
PC=0x411b25
runtime.futex+0x23 /data4/soft/go/go/src/pkg/runtime/linux/amd64/sys.s:137
    runtime.futex()
futexsleep+0x50 /data4/soft/go/go/src/pkg/runtime/linux/thread.c:51
    futexsleep(0x664c18, 0x300000003, 0x0, 0x0)
futexlock+0x85 /data4/soft/go/go/src/pkg/runtime/linux/thread.c:119
    futexlock(0x664c18, 0x100000000)
runtime.notesleep+0x25 /data4/soft/go/go/src/pkg/runtime/linux/thread.c:204
    runtime.notesleep(0x664c18, 0x7fffa0812498)
nextgandunlock+0x146 /data4/soft/go/go/src/pkg/runtime/proc.c:343
    nextgandunlock()
scheduler+0x16f /data4/soft/go/go/src/pkg/runtime/proc.c:536
    scheduler()
runtime.mstart+0x74 /data4/soft/go/go/src/pkg/runtime/proc.c:393
    runtime.mstart()
_rt0_amd64+0x95 /data4/soft/go/go/src/pkg/runtime/amd64/asm.s:69
    _rt0_amd64()
goroutine 2 [3]:
runtime.entersyscall+0x28 /data4/soft/go/go/src/pkg/runtime/proc.c:577
    runtime.entersyscall()
syscall.Syscall6+0x5 /data4/soft/go/go/src/pkg/syscall/asm_linux_amd64.s:40
    syscall.Syscall6()
syscall.EpollWait+0x8d /data4/soft/go/go/src/pkg/syscall/zsyscall_linux_amd64.go:188
    syscall.EpollWait(0x7f6600000006, 0x7f669eb58950, 0x100000001, 0xffffffff, 0xc, ...)
net.*pollster·WaitFD+0xfe /data4/soft/go/go/src/pkg/net/fd_linux.go:116
    net.*pollster·WaitFD(0x7f669eb587a0, 0x0, 0x6400000000, 0x0, 0x0, ...)
net.*pollServer·Run+0xa3 /data4/soft/go/go/src/pkg/net/fd.go:207
    net.*pollServer·Run(0x7f669eb27600, 0x0)
runtime.goexit /data4/soft/go/go/src/pkg/runtime/proc.c:149
    runtime.goexit()
goroutine 1 [4]:
runtime.gosched+0x77 /data4/soft/go/go/src/pkg/runtime/proc.c:558
    runtime.gosched()
runtime.chanrecv+0x18b /data4/soft/go/go/src/pkg/runtime/chan.c:364
    runtime.chanrecv(0x7f669eb3eea0, 0x7f669eb13da8, 0x0, 0x0, 0x0, ...)
runtime.chanrecv1+0x41 /data4/soft/go/go/src/pkg/runtime/chan.c:444
    runtime.chanrecv1(0x7f669eb3eea0, 0x7f669eb123c0)
net.*pollServer·WaitRead+0x52 /data4/soft/go/go/src/pkg/net/fd.go:247
    net.*pollServer·WaitRead(0x7f669eb27600, 0x7f669eb123c0, 0x0, 0x0)
net.*netFD·accept+0x39a /data4/soft/go/go/src/pkg/net/fd.go:579
    net.*netFD·accept(0x7f669eb123c0, 0x43efe8, 0x0, 0x0, 0x0, ...)
net.*TCPListener·AcceptTCP+0x71 /data4/soft/go/go/src/pkg/net/tcpsock.go:261
    net.*TCPListener·AcceptTCP(0x7f669eb0f178, 0x7f669eb13ee0, 0x0, 0x0, 0x1007f6600000001, ...)
net.*TCPListener·Accept+0x49 /data4/soft/go/go/src/pkg/net/tcpsock.go:271
    net.*TCPListener·Accept(0x7f669eb0f178, 0x0, 0x0, 0x0, 0x0, ...)
http.Serve+0x7a /data4/soft/go/go/src/pkg/http/server.go:665
    http.Serve(0x7f669eb27b40, 0x7f669eb0f178, 0x7f669eb4b030, 0x7f669eb0f0a8, 0x0, ...)
main.main+0x886 /home/mc/server/go/tests/setuid2.go:73
    main.main()
runtime.mainstart+0xf /data4/soft/go/go/src/pkg/runtime/amd64/asm.s:77
    runtime.mainstart()
runtime.goexit /data4/soft/go/go/src/pkg/runtime/proc.c:149
    runtime.goexit()
rax     0xfffffffffffffffc
rbx     0x664c18
rcx     0xffffffffffffffff
rdx     0x3
rdi     0x664c18
rsi     0x0
rbp     0x7f669eb13c98
rsp     0x7fffa08123e8
r8      0x0
r9      0x0
r10     0x6601b0
r11     0x206
r12     0x250
r13     0x7fffa0812500
r14     0x0
r15     0x0
rip     0x411b25
rflags  0x206
cs      0x33
fs      0x0
gs      0x0
Trace/breakpoint trap

[rsc]

Comment 16:

Yes, the LockOSThread workaround only applies to one goroutine.

[gopherbot]

Comment 17 by m@capitanio.org:

Thanks, I didn't hit the reload button and missed the comments.
BTW, the crock design is elaborated here ;)
http://www.cs.utexas.edu/~witchel/372/lectures/POSIX_Linux_Threading.pdf

[gopherbot]

Comment 18 by ziutek@Lnet.pl:

I suggest put a clear warning in the header of syscall package about issues that may
cause the use of this package.
Maybe a list of links to known issues for each function would not be a bad idea...

[rsc]

Comment 19:

The syscall package is not for general use.
It has no documentation.  That's not going
to change.  When we have a working Setuid etc
they will be made available as part of package os.
Russ

[alberts]

Comment 20:

Revision a25343ee3016 has added a way of doing Setuid and Setgid as part of
StartProcess, which might help some people.

[rsc]

Comment 21:

Labels changed: added priority-later.

[gopherbot]

Comment 23 by tsar@cosmocode.de:

Is anyone working on that? If there are some experimental patches I would be happy to
test them ;-)
I added a little test case to my own little library to detect non-posix compliant
systems (all the *bsd variants work as expected): https://github.com/sarnowski/mitigation

[rsc]

Comment 24:

No one is working on this as far as I know.
Russ

[rsc]

Comment 25:

Labels changed: added go1.1.

[bpowers]

Comment 26:

I'm looking into this, I have a start to the implementation of what rsc suggested above.
 I've updated the example for go1, along with using the os.Setuid/os.Setgid (an API
change) from my soon-to-be-posted CL.

Attachments:

1. main.go (859 bytes)

[rsc]

Comment 27:

Labels changed: added size-l.

[extemporalgenome]

Comment 28:

The cause of this problem seems closely related to the cause of daemonization problems;
that is, the inability to consistently execute package main code on a vanilla runtime
before any thread spawning occurs (as was trivial with the old init behavior), or
alternatively, the ability to command the runtime to suspend goroutines at their next
non-externally-blocked scheduling point, closing all but one thread (if the runtime can
even temporarily coexist with app code in the main thread). Preemptive scheduling would
certainly make the latter simpler.

[rsc]

Comment 29:

We didn't get to this in time. Going to have to wait for Go 1.2 (or maybe someone will
fix the Linux kernel!).

Labels changed: added go1.2, removed go1.1.

[rsc]

Comment 30:

Labels changed: added go1.3maybe, removed go1.2.

[robpike]

Comment 31:

Labels changed: removed go1.3maybe.

[rsc]

Comment 32:

Labels changed: added go1.3maybe.

[rsc]

Comment 33:

Labels changed: added release-none, removed go1.3maybe.

[rsc]

Comment 34:

Labels changed: added repo-main.

[gopherbot]

Comment 35 by peter@scraperwiki.com:

If it is impossible to get a fix in for this for 1.3, could we at least get a
documentation fix in (at syscall.Setuid et al.)? It would have saved me quite a number
of hours. If so I can submit the change. I'd also be interested at trying to make the
fix if it's not too late for 1.3.

[ianlancetaylor]

Comment 36:

It's certainly not too late for 1.3.  If you want to try to fix this, go for it.  You
might want to send around your proposal for how to fix it first.

[gopherbot]

Comment 37 by digeratus:

Considering all the recent security issues going on, Heartbleed to name just one, I
can't believe this issue is not taking the importance it deserves. Specially, since this
is not even a problem while running on App Engine. Yes, that's right, on App Engine
ListenAndServe is replaced by Google's own version of it. 
http://blog.golang.org/the-app-engine-sdk-and-workspaces-gopath#TOC_2
"The App Engine infrastructure provides its own main function that runs its equivalent
to ListenAndServe. To convert main.go to an App Engine app, drop the call to
ListenAndServe and register the handler in an init function (which runs before main)."
Which makes me wonder what kind of language is Google trying to create here. Is Google's
Golang team in a way saying that to run the language on production you have to do it on
their commercial cloud infrastructure? To think of it, this is something that only comes
from commercially sponsored languages. Python, Perl or any other open source language
that's community spearheaded wouldn't have this type of issue. Who would on the right
mind run an http server as root?
Come on guys, fix this. It's not that hard.

[ianlancetaylor]

Comment 38:

This is only an issue on GNU/Linux.  The right way to handle the problem for a web
server on GNU/Linux is to use setcap.  So this does not seem as urgent to me as it
apparently does to you.
That said, of course it would be good to fix this.  But I think it is harder than you
think it is.  I would be glad to be proven wrong.

Labels changed: added suggested, os-linux, removed priority-later.

[gopherbot]

Comment 39 by digeratus:

With all due respect, but if you don't consider GNU/Linux to be an important web server
platform, I don't know what is. At the same time, how hard is it to do what requires
"root" permission and the lower it to a lower permission user?

[gopherbot]

Comment 40 by digeratus:

Take a look at how NGINX handles this, line 36:
http://trac.nginx.org/nginx/browser/nginx/src/os/unix/ngx_daemon.c#L36

[gopherbot]

Comment 41 by peter.waller:

digeratus, unfortunately, it is harder than you think to pull off setuid/setsid
correctly. And as you point out, it is security sensitive, so it is important to get
correct.
You can find some discussion here:
https://groups.google.com/forum/#!searchin/golang-dev/setuid$20implementing/golang-dev/sVCQl9adDgg/-EeOdMaWT48J
Comparing to nginx is unhelpful, since it runs under a completely different runtime. A
more apt comparison would be how glibc handles it:
https://github.com/lattera/glibc/search?q=SIGSETXID&ref=cmdform
Which involves sending and receiving an SIGSETXID signal and synchronizing all of the
threads.
I've had your exact frustration with this issue. It seems simple, but it isn't, sadly.
i@golang.org: could you elaborate on the setcap solution?

[minux]

Comment 42:

listening on privilege ports on linux needs not the full root, but only
a single capabilities(8) called CAP_NET_BIND_SERVICE.
you can grant any capability on any executable file by using the setcap(8)
command.
e.g.:
sudo setcap cap_net_bind_service=+ep go_server
and then you can run go_server as a normal user (e.g. nobody).

[gopherbot]

Comment 43 by peter.waller:

Aha! I was expecting a syscall, not an extended file attribute, which is why I was
confused when I looked at the setcap page. Awesome, thanks minux.

[ianlancetaylor]

Comment 44:

> With all due respect, but if you don't consider GNU/Linux to be an important web
server platform, I don't know what is.
I don't know what you mean by that.  Of course GNU/Linux is an important web server
platform.
> At the same time, how hard is it to do what requires "root" permission and the lower
it to a lower permission user?
On GNU/Linux, it's best to use the setcap mechanism, and not use root permission at all.
 setcap has been around since Linux 2.2, it's not some newfangled thing.
That said, we all agree that this bug should be fixed.  It would be great if you worked
on fixing it, rather than simply telling us that it is easy to fix.  Thanks.

[gopherbot]

Comment 45 by StevensElectronicMail:

Given that setuid is broken with Go can it be removed (or maybe somehow tagged with a
warning) while we're waiting for a fix so that nobody uses it by accident. I know this
would break backwards compatibility but in this case it SHOULD break backwards
compatibility because code that uses it most probably has a security hole.

[davecheney]

Comment 46:

I don't know about removing the symbol, but I think there is an argument to be made that
it should _always_ return an error on systems where it is known to be faulty.

[ianlancetaylor]

Comment 47:

SGTM

[gopherbot]

Comment 48:

CL https://golang.org/cl/106170043 mentions this issue.

[davecheney]

Comment 49:

This issue was updated by revision 343b4ba8c1ad8a29b6dd19cb101273b57a26c9b.

This proposal disables Setuid and Setgid on all linux platforms.
issue #1435 has been open for a long time, and it is unlikely to be addressed soon so an
argument was made by a commenter
https://golang.org/issue/1435?c=45
That these functions should made to fail rather than succeed in their broken state.
LGTM=ruiu, iant
R=iant, ruiu
CC=golang-codereviews
https://golang.org/cl/106170043

[rui314]

Comment 50:

Can we fix this before Go 1.4 by any chance? I read the glibc code briefly. It looks
like we need to (1) set a signal handler for SIGSETXID that calls setuid/setgid system
call from the handler, and (2) signal all threads with SIGSETXID when setuid/setgid Go
function is called. Because we don't provide syscall.sigaction, I think we don't need to
worry about the case that a user accidentally disable SIGSETXID system call. Is there
anything other than that we need to do for this?

[ianlancetaylor]

Comment 51:

Since Setuid is used for security, and is often used to drop privileges, it's fairly
important to get it right.
In a pure Go program with no cgo, I believe we can freeze the world and force each
thread to use the system call.  I don't think we need any signals in that case.
The problem is a Go program that uses cgo.  We have to assume that cgo code has created
threads that we don't know about.  We also have to assume that cgo code can call setuid.
 Because of that possibility, I don't think we can provide a SIGSETXID signal handler
ourselves for a cgo program; see issue #3871.
Adding this all together, it now occurs to me that it may work to make syscall.Setuid a
function that uses cgo to call the C setuid function.  That will force the use of cgo
for any program that uses syscall.Setuid, meaning that every thread will be created
using pthread_create and that we should get the C library signal handler for SIGSETXID.
Want to see if that will work?