Skip to contentSkip to conversation

fs/cpio/init: unbreak ttyname_r() on glibc after dropping /dev/console exec

PATCH New Peter Korsgaard 2 participants 5 messages Aug 29, 2020

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 fs/cpio/init: unbreak ttyname_r() on glibc after dropping /dev/console exec

Commit 98a6f1fc02e41 (fs/cpio: make initramfs init script survive 'console=' kernel argument) dropped the explicit /dev/console execs for fd 0,1,2, as they fail when booted with console= and aren't really needed as the kernel will setup fd 0,1,2 from /dev/console before executing the initramfs anyway.

Not doing this unfortunately confuses glibc's ttyname_r(3) implementation (used by E.G. busybox/coreutils 'tty'), causing it to fail with ENOENT as it does a fstat on fd 0 and tries to match up st_ino / st_dev against the entries in /dev (since glibc 2.26):

commit 15e9a4f378c8607c2ae1aa465436af4321db0e23
Author: Christian Brauner <[email protected]>
Date: Fri Jan 27 15:59:59 2017 +0100

linux ttyname and ttyname_r: do not return wrong results

If a link (say /proc/self/fd/0) pointing to a device, say /dev/pts/2, in a parent mount namespace is passed to ttyname, and a /dev/pts/2 exists (in a different devpts) in the current namespace, then it returns /dev/pts/2. But /dev/pts/2 is NOT the current tty, it is a different file and device.

Detect this case and return ENODEV. Userspace can choose to take this as a hint that the fd points to a tty device but to act on the fd rather than the link.

The reason it fails is that we manually mount devtmpfs on /dev in /init, so the /dev/console used by the kernel (in rootfs) is not the same file as /dev/console at runtime (in devtmpfs).

Notice: Once logged in, tty does work correctly. Presumably login reopens stdin/stdout/stderr.

To fix this, re-add the exec of /dev/console for fd 0,1,2, but only do so if possible. Because of the above mentioned shell behaviour (specified by POSIX [0]), perform this check in a subshell.

[0] https://pubs.opengroup.org/onlinepubs/9699919799/utilities/V3_chap02.html#tag_18_20_01

---
 fs/cpio/init | 11 +++++++++++
 1 file changed, 11 insertions(+)
fs/cpio/init +11 -0
index b0af18b67a..f74ef7e15f 100755
@@ -1,4 +1,15 @@
 #!/bin/sh
 # devtmpfs does not get automounted for initramfs
 /bin/mount -t devtmpfs devtmpfs /dev
+
+# use the /dev/console device node from devtmpfs if possible to not
+# confuse glibc's ttyname_r().
+# This may fail (E.G. booted with console=), and errors from exec will
+# terminate the shell, so use a subshell for the test
+if (exec 0</dev/console) 2>/dev/null; then
+    exec 0</dev/console
+    exec 1>/dev/console
+    exec 2>/dev/console
+fi
+
 exec /sbin/init "$@"
Signature · 1 line
2.20.1
#1 lore raw

PATCH 1/1 fs/cpio/init: unbreak ttyname_r() on glibc after dropping /dev/console exec

Peter, All,

On 2020-08-29 15:09 +0200, Peter Korsgaard spake thusly:

Quoted 53 lines

Commit 98a6f1fc02e41 (fs/cpio: make initramfs init script survive 'console=' kernel argument) dropped the explicit /dev/console execs for fd 0,1,2, as they fail when booted with console= and aren't really needed as the kernel will setup fd 0,1,2 from /dev/console before executing the initramfs anyway.

Not doing this unfortunately confuses glibc's ttyname_r(3) implementation (used by E.G. busybox/coreutils 'tty'), causing it to fail with ENOENT as it does a fstat on fd 0 and tries to match up st_ino / st_dev against the entries in /dev (since glibc 2.26):

commit 15e9a4f378c8607c2ae1aa465436af4321db0e23
Author: Christian Brauner <[email protected]>
Date: Fri Jan 27 15:59:59 2017 +0100

linux ttyname and ttyname_r: do not return wrong results

If a link (say /proc/self/fd/0) pointing to a device, say /dev/pts/2, in a parent mount namespace is passed to ttyname, and a /dev/pts/2 exists (in a different devpts) in the current namespace, then it returns /dev/pts/2. But /dev/pts/2 is NOT the current tty, it is a different file and device.

Detect this case and return ENODEV. Userspace can choose to take this as a hint that the fd points to a tty device but to act on the fd rather than the link.

The reason it fails is that we manually mount devtmpfs on /dev in /init, so the /dev/console used by the kernel (in rootfs) is not the same file as /dev/console at runtime (in devtmpfs).

Notice: Once logged in, tty does work correctly. Presumably login reopens stdin/stdout/stderr.

To fix this, re-add the exec of /dev/console for fd 0,1,2, but only do so if possible. Because of the above mentioned shell behaviour (specified by POSIX [0]), perform this check in a subshell.

[0] https://pubs.opengroup.org/onlinepubs/9699919799/utilities/V3_chap02.html#tag_18_20_01

---
 fs/cpio/init | 11 +++++++++++
 1 file changed, 11 insertions(+)
fs/cpio/init +0 -0
index b0af18b67a..f74ef7e15f 100755
@@ -1,4 +1,15 @@
 #!/bin/sh
 # devtmpfs does not get automounted for initramfs
 /bin/mount -t devtmpfs devtmpfs /dev

fs/cpio/init line 13

Quoted 10 lines
fs/cpio/init +10 -0
+
+# use the /dev/console device node from devtmpfs if possible to not
+# confuse glibc's ttyname_r().
+# This may fail (E.G. booted with console=), and errors from exec will
+# terminate the shell, so use a subshell for the test
+if (exec 0</dev/console) 2>/dev/null; then
+    exec 0</dev/console
+    exec 1>/dev/console
+    exec 2>/dev/console
+fi

You forgot the else part, which should redirect to /dev/null, so that we do have valid fds in the case /dev/console does not really exist, no?

Regards,
Yann E. MORIN.

Quoted 4 lines
fs/cpio/init +0 -0
 exec /sbin/init "$@"
Signature · 1 line
2.20.1
Signature · 6 lines
.-----------------.--------------------.------------------.--------------------.
|  Yann E. MORIN  | Real-Time Embedded | /"\ ASCII RIBBON | Erics' conspiracy: |
| +33 662 376 056 | Software  Designer | \ / CAMPAIGN     |  ___               |
| +33 561 099 427 `------------.-------:  X  AGAINST      |  \e/  There is no  |
| http://ymorin.is-a-geek.org/ | _/*\_ | / \ HTML MAIL    |   v   conspiracy.  |
'------------------------------^-------^------------------^--------------------'
#2 lore raw

PATCH 1/1 fs/cpio/init: unbreak ttyname_r() on glibc after dropping /dev/console exec

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

"Yann" == Yann E MORIN <[email protected]> writes:

Hi,

Quoted 4 lines
Quoted 4 lines
fs/cpio/init +4 -0
+# use the /dev/console device node from devtmpfs if possible to not
+# confuse glibc's ttyname_r().
+# This may fail (E.G. booted with console=), and errors from exec will
+# terminate the shell, so use a subshell for the test
Quoted 8 lines

fs/cpio/init line 13

Quoted 5 lines
fs/cpio/init +5 -0
+if (exec 0</dev/console) 2>/dev/null; then
+    exec 0</dev/console
+    exec 1>/dev/console
+    exec 2>/dev/console
+fi

You forgot the else part, which should redirect to /dev/null, so that we do have valid fds in the case /dev/console does not really exist, no?

Well, the kernel already takes care of that since commit 2bd3a997befc2 (E.G. as explained by buildroot commit 98a6f1fc02e41), so I am not sure what explicitly reopening /dev/null would achive?

ttyname_r() / tty already returns ENOENT (not a tty) for this case without reopening /dev/null:

tty </dev/null
not a tty

Signature · 1 line
Bye, Peter Korsgaard
#3 lore raw

PATCH 1/1 fs/cpio/init: unbreak ttyname_r() on glibc after dropping /dev/console exec

In reply to Peter Korsgaard

Peter, All,

On 2020-08-29 15:09 +0200, Peter Korsgaard spake thusly:

Quoted 34 lines

Commit 98a6f1fc02e41 (fs/cpio: make initramfs init script survive 'console=' kernel argument) dropped the explicit /dev/console execs for fd 0,1,2, as they fail when booted with console= and aren't really needed as the kernel will setup fd 0,1,2 from /dev/console before executing the initramfs anyway.

Not doing this unfortunately confuses glibc's ttyname_r(3) implementation (used by E.G. busybox/coreutils 'tty'), causing it to fail with ENOENT as it does a fstat on fd 0 and tries to match up st_ino / st_dev against the entries in /dev (since glibc 2.26):

commit 15e9a4f378c8607c2ae1aa465436af4321db0e23
Author: Christian Brauner <[email protected]>
Date: Fri Jan 27 15:59:59 2017 +0100

linux ttyname and ttyname_r: do not return wrong results

If a link (say /proc/self/fd/0) pointing to a device, say /dev/pts/2, in a parent mount namespace is passed to ttyname, and a /dev/pts/2 exists (in a different devpts) in the current namespace, then it returns /dev/pts/2. But /dev/pts/2 is NOT the current tty, it is a different file and device.

Detect this case and return ENODEV. Userspace can choose to take this as a hint that the fd points to a tty device but to act on the fd rather than the link.

The reason it fails is that we manually mount devtmpfs on /dev in /init, so the /dev/console used by the kernel (in rootfs) is not the same file as /dev/console at runtime (in devtmpfs).

Notice: Once logged in, tty does work correctly. Presumably login reopens stdin/stdout/stderr.

Quoted 7 lines

To fix this, re-add the exec of /dev/console for fd 0,1,2, but only do so if possible. Because of the above mentioned shell behaviour (specified by POSIX [0]), perform this check in a subshell.

[0] https://pubs.opengroup.org/onlinepubs/9699919799/utilities/V3_chap02.html#tag_18_20_01

Applied to master, thanks.

Regards,
Yann E. MORIN.

Quoted 27 lines
---
 fs/cpio/init | 11 +++++++++++
 1 file changed, 11 insertions(+)
fs/cpio/init +11 -0
index b0af18b67a..f74ef7e15f 100755
@@ -1,4 +1,15 @@
 #!/bin/sh
 # devtmpfs does not get automounted for initramfs
 /bin/mount -t devtmpfs devtmpfs /dev
+
+# use the /dev/console device node from devtmpfs if possible to not
+# confuse glibc's ttyname_r().
+# This may fail (E.G. booted with console=), and errors from exec will
+# terminate the shell, so use a subshell for the test
+if (exec 0</dev/console) 2>/dev/null; then
+    exec 0</dev/console
+    exec 1>/dev/console
+    exec 2>/dev/console
+fi
+
 exec /sbin/init "$@"
Signature · 1 line
2.20.1
Signature · 6 lines
.-----------------.--------------------.------------------.--------------------.
|  Yann E. MORIN  | Real-Time Embedded | /"\ ASCII RIBBON | Erics' conspiracy: |
| +33 662 376 056 | Software  Designer | \ / CAMPAIGN     |  ___               |
| +33 561 099 427 `------------.-------:  X  AGAINST      |  \e/  There is no  |
| http://ymorin.is-a-geek.org/ | _/*\_ | / \ HTML MAIL    |   v   conspiracy.  |
'------------------------------^-------^------------------^--------------------'
#4 lore raw

PATCH 1/1 fs/cpio/init: unbreak ttyname_r() on glibc after dropping /dev/console exec

In reply to Peter Korsgaard
Quoted 36 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

"Peter" == Peter Korsgaard <[email protected]> writes:

Commit 98a6f1fc02e41 (fs/cpio: make initramfs init script survive 'console=' kernel argument) dropped the explicit /dev/console execs for fd 0,1,2, as they fail when booted with console= and aren't really needed as the kernel will setup fd 0,1,2 from /dev/console before executing the initramfs anyway.

Not doing this unfortunately confuses glibc's ttyname_r(3) implementation (used by E.G. busybox/coreutils 'tty'), causing it to fail with ENOENT as it does a fstat on fd 0 and tries to match up st_ino / st_dev against the entries in /dev (since glibc 2.26):

commit 15e9a4f378c8607c2ae1aa465436af4321db0e23
Author: Christian Brauner <[email protected]>
Date: Fri Jan 27 15:59:59 2017 +0100

linux ttyname and ttyname_r: do not return wrong results

If a link (say /proc/self/fd/0) pointing to a device, say /dev/pts/2, in a parent mount namespace is passed to ttyname, and a /dev/pts/2 exists (in a different devpts) in the current namespace, then it returns /dev/pts/2. But /dev/pts/2 is NOT the current tty, it is a different file and device.

Detect this case and return ENODEV. Userspace can choose to take this as a hint that the fd points to a tty device but to act on the fd rather than the link.

The reason it fails is that we manually mount devtmpfs on /dev in /init, so the /dev/console used by the kernel (in rootfs) is not the same file as /dev/console at runtime (in devtmpfs).

Notice: Once logged in, tty does work correctly. Presumably login reopens stdin/stdout/stderr.

Quoted 7 lines

To fix this, re-add the exec of /dev/console for fd 0,1,2, but only do so if possible. Because of the above mentioned shell behaviour (specified by POSIX [0]), perform this check in a subshell.

[0] https://pubs.opengroup.org/onlinepubs/9699919799/utilities/V3_chap02.html#tag_18_20_01

Committed to 2020.02.x and 2020.05.x, thanks.

Signature · 1 line
Bye, Peter Korsgaard