Skip to contentSkip to conversation

possible regression with pty.c commit

DISCUSSION Sergey Senozhatsky 5 participants 15 messages Jul 11, 2009 – Jul 19, 2009

lore
Share & tools
lore
#0 lore raw

Hello,
commit d945cb9cce20ac7143c2de8d88b187f62db99bdc ("pty: Rework the pty layer to use the normal buffering logic") seems to brake kdesu.

(quote: "Because some su implementations (i.e. the one from Red HatВ®) don't want to read the password from stdin, KDE su creates a pty/tty pair and executes su with it's standard filedescriptors connected to the tty.")

Revert "pty: Rework the pty layer to use the normal buffering logic" (commit d945cb9cce20ac7143c2de8d88b187f62db99bdc) makes kdesu work again.

strace /usr/bin/kdesu -u root -c top:
//tons of kde stuff were cut

execve("/usr/bin/kdesu", ["/usr/bin/kdesu", "-u", "root", "-c", "top"], [/* 36 vars */]) = 0
brk(0) = 0x9268000
...

sigreturn()                             = ? (mask now [])
access("/bin/su", X_OK)                 = 0
open("/dev/ptmx", O_RDWR)               = 10

ioctl(10, SNDCTL_TMR_TIMEBASE or TCGETS, {B38400 opost isig icanon echo ...}) = 0
ioctl(10, TIOCGPTN, [10]) = 0
stat64("/dev/pts/10", {st_mode=S_IFCHR|0620, st_rdev=makedev(136, 10), ...}) = 0
ioctl(10, SNDCTL_TMR_TIMEBASE or TCGETS, {B38400 opost isig icanon echo ...}) = 0
ioctl(10, TIOCGPTN, [10]) = 0
stat64("/dev/pts/10", {st_mode=S_IFCHR|0620, st_rdev=makedev(136, 10), ...}) = 0
statfs("/dev/pts/10", {f_type="DEVPTS_SUPER_MAGIC", f_bsize=4096, f_blocks=0, f_bfree=0, f_bavail=0, f_files=0, f_ffree=0, f_fsid={0, 0}, f_namelen=255, f_frsize=4096}) = 0
ioctl(10, TIOCSPTLCK, [0]) = 0
open("/dev/pts/10", O_RDWR) = 11
clone(child_stack=0, flags=CLONE_CHILD_CLEARTID|CLONE_CHILD_SETTID|SIGCHLD, child_tidptr=0xb68c5718) = 4072

close(11)                               = 0
fcntl64(10, F_GETFL)                    = 0x2 (flags O_RDWR)
read(10, "Password: ", 255)             = 10
select(11, [10], NULL, NULL, {0, 100000}) = 0 (Timeout)
kill(4072, SIG_0)                       = 0
open("/dev/pts/10", O_RDWR)             = 11
kill(4072, SIG_0)                       = 0

ioctl(11, SNDCTL_TMR_TIMEBASE or TCGETS, {B38400 -opost isig icanon -echo ...}) = 0
close(11) = 0
kill(4072, SIG_0) = 0
write(10, "**~WRITE_PASSWORD~**", 11) = 11

write(10, "\n", 1)                      = 1
fcntl64(10, F_GETFL)                    = 0x2 (flags O_RDWR)
read(10, "\n", 255)                     = 1
fcntl64(10, F_GETFL)                    = 0x2 (flags O_RDWR)
read(10, "kdesu_stub\n", 255)           = 11
open("/dev/pts/10", O_RDWR)             = 11

ioctl(11, SNDCTL_TMR_TIMEBASE or TCGETS, {B38400 -opost isig icanon echo ...}) = 0 ioctl(11, SNDCTL_TMR_TIMEBASE or TCGETS, {B38400 -opost isig icanon echo ...}) = 0 ioctl(11, SNDCTL_TMR_START or TCSETS, {B38400 -opost isig icanon -echo ...}) = 0 ioctl(11, SNDCTL_TMR_TIMEBASE or TCGETS, {B38400 -opost isig icanon -echo ...}) = 0

close(11)                               = 0
write(10, "stop", 4)                    = 4
write(10, "\n", 1)                      = 1
fcntl64(10, F_GETFL)                    = 0x2 (flags O_RDWR)
read(10, 0xbf840858, 255)               = -1 EIO (Input/output error)		<<<<<<<<<

--- SIGCHLD (Child exited) @ 0 (0) ---
write(5, "\0", 1) = 1
sigreturn() = ? (mask now [])
.....

As for "bug 13522: BUG: scheduling while atomic" - I'll test it.

Sergey

#1 lore raw
Quoted 3 lines · On Sat, 11 Jul 2009 22:15:56 +0300 Sergey Senozhatsky…

Hello,
commit d945cb9cce20ac7143c2de8d88b187f62db99bdc ("pty: Rework the pty layer to use the normal buffering logic") seems to brake kdesu.

This looks like a timing bug in kdesu at first glance but it may be more complex.

Quoted 1 line

close(11) = 0

We close one side of the pty/tty pair

Quoted 4 lines
write(10, "stop", 4)                    = 4
write(10, "\n", 1)                      = 1
fcntl64(10, F_GETFL)                    = 0x2 (flags O_RDWR)
read(10, 0xbf840858, 255)               = -1 EIO (Input/output error)		<<<<<<<<<

At this point the other side is closed, we have a hangup and the read correctly I think gets -EIO.

I will have a look at kdesu on Monday, I've got Fedora setups here so hopefully I can reproduce it simply.

#2 lore raw
Quoted 19 lines · Alan Cox
Quoted 3 lines · On Sat, 11 Jul 2009 22:15:56 +0300 Sergey Senozhatsky…

Hello,
commit d945cb9cce20ac7143c2de8d88b187f62db99bdc ("pty: Rework the pty layer to use the normal buffering logic") seems to brake kdesu.

This looks like a timing bug in kdesu at first glance but it may be more complex.

Quoted 1 line

close(11) = 0

We close one side of the pty/tty pair

Quoted 4 lines
write(10, "stop", 4)                    = 4
write(10, "\n", 1)                      = 1
fcntl64(10, F_GETFL)                    = 0x2 (flags O_RDWR)
read(10, 0xbf840858, 255)               = -1 EIO (Input/output error)		<<<<<<<<<
Quoted 6 lines

At this point the other side is closed, we have a hangup and the read correctly I think gets -EIO.

I will have a look at kdesu on Monday, I've got Fedora setups here so hopefully I can reproduce it simply.

Alan, I forgot to tell - I'm using KDE 3.5.9 (3.5.10). Don't know whether this can be reproduced with KDE 4.x.x.

Sergey

#3 lore raw
Quoted 1 line

Alan, I forgot to tell - I'm using KDE 3.5.9 (3.5.10). Don't know whether this can be reproduced with KDE 4.x.x.

Dumping out the traces tty side the tty code appears to be working correctly. The userspace on the other hand appears broken and to only work by chance with the old code.

For one it wants for new output to appear before checking for "Password:" and in doing so consumes any partial output it receives without checking. I suspect it should be doing line += more

One for the KDE people as its not the kind of thing we can fudge back kernel side if my diagnosis is right.

#4 lore raw
Quoted 6 lines · Alan Cox
Quoted 1 line

Alan, I forgot to tell - I'm using KDE 3.5.9 (3.5.10). Don't know whether this can be reproduced with KDE 4.x.x.

Dumping out the traces tty side the tty code appears to be working correctly. The userspace on the other hand appears broken and to only work by chance with the old code.

Quoted 7 lines

For one it wants for new output to appear before checking for "Password:" and in doing so consumes any partial output it receives without checking. I suspect it should be doing line += more

One for the KDE people as its not the kind of thing we can fudge back kernel side if my diagnosis is right.

Hello,
Sorry for delay. d945cb9cce20ac7143c2de8d88b187f62db99bdc is not the first bad commit as I've managed to make kdesu work simply adding some senseless instructions to pty_write_room(). Anyway.

As for kdesu, it's obvious that kernel should be free from code that simply makes some buggy user space programm working (even one that was working for years until 2.6.31).
Lots of versions could be affected (4.a.b, 3.c.d, ...) and I'm just not sure that all of them will be updated ( not all, but popular ones. like 3.5.x ). We'll see.

Thanks,
Sergey

#5 lore raw
In reply to Alan Cox

Hello Alan,
I'm having another problem which I guess may be caused by pty/tty changes. The problem is that ppp connection constantly hangups under load (downloading) (it works perfectly with 30 kernel).

syslog:
pptp[1942]: anon log[logecho:pptp_ctrl.c:677]: Echo Reply received. pptp[1942]: anon log[logecho:pptp_ctrl.c:677]: Echo Reply received. pptp[1942]: anon log[logecho:pptp_ctrl.c:677]: Echo Reply received. pptp[1942]: anon log[logecho:pptp_ctrl.c:677]: Echo Reply received. pptp[1942]: anon log[logecho:pptp_ctrl.c:677]: Echo Reply received. pptp[1942]: anon log[logecho:pptp_ctrl.c:677]: Echo Reply received. pptp[1942]: anon log[logecho:pptp_ctrl.c:677]: Echo Reply received. pptp[1942]: anon log[logecho:pptp_ctrl.c:677]: Echo Request received. pptp[1942]: anon log[ctrlp_rep:pptp_ctrl.c:251]: Sent control packet type is 6 'Echo-Reply' pppd[1929]: No response to 4 echo-requests
pppd[1929]: Serial link appears to be disconnected.
pppd[1929]: Connect time 8.5 minutes.

In average it works ~10 minutes.
I did "strace -ff -F -tt -s 200 -o ... pon ..." which produced 11MiB and 46 files (it'll take some time to dig). Do you have any ability to test ppp under load?

Sergey

#6 lore raw
Quoted 19 lines · Sergey Senozhatsky

Hello Alan,
I'm having another problem which I guess may be caused by pty/tty changes. The problem is that ppp connection constantly hangups under load (downloading) (it works perfectly with 30 kernel).

syslog:
pptp[1942]: anon log[logecho:pptp_ctrl.c:677]: Echo Reply received. pptp[1942]: anon log[logecho:pptp_ctrl.c:677]: Echo Reply received. pptp[1942]: anon log[logecho:pptp_ctrl.c:677]: Echo Reply received. pptp[1942]: anon log[logecho:pptp_ctrl.c:677]: Echo Reply received. pptp[1942]: anon log[logecho:pptp_ctrl.c:677]: Echo Reply received. pptp[1942]: anon log[logecho:pptp_ctrl.c:677]: Echo Reply received. pptp[1942]: anon log[logecho:pptp_ctrl.c:677]: Echo Reply received. pptp[1942]: anon log[logecho:pptp_ctrl.c:677]: Echo Request received. pptp[1942]: anon log[ctrlp_rep:pptp_ctrl.c:251]: Sent control packet type is 6 'Echo-Reply' pppd[1929]: No response to 4 echo-requests
pppd[1929]: Serial link appears to be disconnected.
pppd[1929]: Connect time 8.5 minutes.

Quoted 5 lines

In average it works ~10 minutes.
I did "strace -ff -F -tt -s 200 -o ... pon ..." which produced 11MiB and 46 files (it'll take some time to dig). Do you have any ability to test ppp under load?

Sergey

Hello everyone,

I can reproduce this in mere seconds, using an otherwise idle ppp link and ping -f -s256 -l256.
Usually, the first batch of packets hangs the pty, but if the link is not entirely idle, the ping may work for a few more seconds. In fact, bisect points to commit d945cb9cce20ac7143c2de8d88b187f62db99bdc ("pty:
Rework the pty layer to use the normal buffering logic").

Regards,
Pierre

#7 lore raw
In reply to Sergey Senozhatsky
Quoted 4 lines · On Tue, 14 Jul 2009 19:28:37 +0300 Sergey Senozhatsky…

Hello Alan,
I'm having another problem which I guess may be caused by pty/tty changes. The problem is that ppp connection constantly hangups under load (downloading) (it works perfectly with 30 kernel).

Have you also got the unthrottle change reverted - without that revert it certainly will hang as you describe.

#8 lore raw
Quoted 8 lines · Alan Cox
Quoted 4 lines

Hello Alan,
I'm having another problem which I guess may be caused by pty/tty changes. The problem is that ppp connection constantly hangups under load (downloading) (it works perfectly with 30 kernel).

Have you also got the unthrottle change reverted - without that revert it certainly will hang as you describe.

Hello Alan, do yo mean

drivers/net/ppp_async.c
commit a6540f731d506d9e82444cf0020e716613d4c46c
ppp: Fix throttling bugs

- tty_unthrottle(tty);

I'll revert it and test.

Thanks,
Sergey

#9 lore raw
Quoted 12 lines · On Wed, 15 Jul 2009 13:27:33 +0300 Sergey Senozhatsky…
Quoted 8 lines · Alan Cox
Quoted 4 lines

Hello Alan,
I'm having another problem which I guess may be caused by pty/tty changes. The problem is that ppp connection constantly hangups under load (downloading) (it works perfectly with 30 kernel).

Have you also got the unthrottle change reverted - without that revert it certainly will hang as you describe.

Hello Alan, do yo mean

Quoted 7 lines

drivers/net/ppp_async.c
commit a6540f731d506d9e82444cf0020e716613d4c46c
ppp: Fix throttling bugs

- tty_unthrottle(tty);

I'll revert it and test.

I do yes

#10 lore raw
Quoted 6 lines · Alan Cox
Quoted 6 lines · Alan Cox
Quoted 5 lines
Quoted 4 lines

Hello Alan,
I'm having another problem which I guess may be caused by pty/tty changes. The problem is that ppp connection constantly hangups under load (downloading) (it works perfectly with 30 kernel).

Quoted 6 lines
Quoted 4 lines
Quoted 3 lines

Have you also got the unthrottle change reverted - without that revert it certainly will hang as you describe.

I do yes

d945cb9cce20ac7143c2de8d88b187f62db99bdc "pty: Rework the pty layer to use the normal buffering logic" and REVERTED a6540f731d506d9e82444cf0020e716613d4c46c "ppp: Fix throttling bugs"

seems to work.

Sergey

#11 lore raw
Quoted 11 lines · Sergey Senozhatsky
Quoted 10 lines · Alan Cox
Quoted 10 lines · Alan Cox
Quoted 8 lines
Quoted 4 lines

Hello Alan,
I'm having another problem which I guess may be caused by pty/tty changes. The problem is that ppp connection constantly hangups under load (downloading) (it works perfectly with 30 kernel).

Have you also got the unthrottle change reverted - without that revert it certainly will hang as you describe.

Quoted 7 lines
Quoted 2 lines

I do yes

 d945cb9cce20ac7143c2de8d88b187f62db99bdc "pty: Rework the pty layer to use  the normal buffering logic" and REVERTED a6540f731d506d9e82444cf0020e716613d4c46c "ppp: Fix throttling bugs"

seems to work.

I have same result, the pppoe connection is unstable with the patch. Are there better solution for this issue?

Quoted 14 lines

       Sergey

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.9 (GNU/Linux)

iJwEAQECAAYFAkpeC/cACgkQfKHnntdSXjSGcwQAuDgNU1r+1lgdto4HdBaejLW8 +DbEVuixCG4PPBM9nlCxI2nK0KydKF/8KRw5iI/GszXlk4WXhlwDcWbAK7BoUPyH 3D4uw17BtcYbtn2cmPFeAC4pMVbrDr3CsstBNW6F7K37VYL8bgIfQrEoL9Oi5SUE tNV6lyNnuGq/3zSqRA8=
=ROb6
-----END PGP SIGNATURE-----

Signature · 2 lines
Regards
dave
#12 lore raw
Quoted 8 lines · Dave Young
Quoted 4 lines

 d945cb9cce20ac7143c2de8d88b187f62db99bdc "pty: Rework the pty layer to use  the normal buffering logic" and REVERTED a6540f731d506d9e82444cf0020e716613d4c46c "ppp: Fix throttling bugs"

seems to work.

I have same result, the pppoe connection is unstable with the patch. Are there better solution for this issue?

Hello Dave.
The solution is in rc3-git2.
Linus has reverted "throttling" commits.

git log drivers/net/ppp_async.c
git log drivers/net/ppp_synctty.c
4a21b8cb3550f19f838f7c48345fbbf6a0e8536b
Revert "ppp: Fix throttling bugs"

Sergey

#13 lore raw
Quoted 19 lines · Sergey Senozhatsky
Quoted 8 lines · Dave Young
Quoted 4 lines

 d945cb9cce20ac7143c2de8d88b187f62db99bdc "pty: Rework the pty layer to use  the normal buffering logic" and REVERTED a6540f731d506d9e82444cf0020e716613d4c46c "ppp: Fix throttling bugs"

seems to work.

I have same result, the pppoe connection is unstable with the patch. Are there better solution for this issue?

Hello Dave.
The solution is in rc3-git2.
Linus has reverted "throttling" commits.

git log drivers/net/ppp_async.c
git log drivers/net/ppp_synctty.c
4a21b8cb3550f19f838f7c48345fbbf6a0e8536b
Revert "ppp: Fix throttling bugs"

Quoted 5 lines

       Sergey

Thanks for tell, and then the "scheduling in atomic" problem?

Quoted 11 lines

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.9 (GNU/Linux)

iJwEAQECAAYFAkpgN5gACgkQfKHnntdSXjRg0wQAhef6hdvDaOXuQSsxFLH/l9H1 99NtAA8ewJ7Rx6/lNkzLu9n/A8qJtdoZxNP9xQDeBbeMXaH+NA5+FSo48mGbb/jp ePjb3vIIOzRRqGWsKYFdn0m4JW5Bw/2IvVFbHBPfa9C+ZDBPodiyDHI9gZi8bF99 xc5uyDLAlCksK+WYu+s=
=BLJz
-----END PGP SIGNATURE-----

Signature · 2 lines
Regards
dave
#14 lore raw
In reply to Alan Cox

Here is a testcase with expect:

#!/usr/bin/expect -f
spawn echo foo bar

expect {
    "foo bar" {puts $expect_out(buffer)}
}

With d945cb9cce20ac7143c2de8d88b187f62db99bdc reverted, I get the expected output:

$ expect -f pty.exp
spawn echo foo bar
foo bar
foo bar

Without the revert the expect process receives EIO when it tries to read the output of the echo process which already died.

Andreas.

Signature · 3 lines
Andreas Schwab, [email protected]
GPG Key fingerprint = 58CA 54C7 6D53 942B 1756  01D3 44D5 214B 8276 4ED5
"And now for something completely different."