tty: tty_io: update timestamps on all device nodes
Subsystems (2)
PATCH 1/2 tty: tty_io: update timestamps on all device nodes
User space applications watch for timestamp changes on character device files in order to determine idle time of a given terminal session. For example, "w" program uses this information to populate the IDLE column of its output [1]. Similarly, systemd-logind has optional feature where it uses atime of the tty character device to determine if there was activity on the terminal associated with the logind's session object. If there was no activity for a configured period of time then logind will terminate such session [2].
Now, usually (e.g. bash running on the terminal) the use of the terminal will update timestamps (atime and mtime) on the corresponding terminal character device. However, if access to the terminal, e.g. /dev/pts/0, is performed through magic character device /dev/tty then such access obviously changes the state of the terminal, however timestamps on the device that correspond to the terminal (/dev/pts/0) are not updated.
This patch makes sure that we update timestamps on *all* character devices that correspond to the given tty, because outside observers (w, systemd-logind) are maybe checking these timestamps. Obviously, they can not check timestamps on /dev/tty as that has per-process meaning.
[1] https://gitlab.com/procps-ng/procps/-/blob/v4.0.0/w.c#L286
[2] https://github.com/systemd/systemd/blob/v252/NEWS#L477
--- drivers/tty/tty_io.c | 32 +++++++++++++++++++++----------- 1 file changed, 21 insertions(+), 11 deletions(-)
drivers/tty/tty_io.c +21 -11
@@ -101,6 +101,7 @@#include <linux/compat.h> #include <linux/uaccess.h> #include <linux/termios_internal.h> +#include <linux/fs.h> #include <linux/kbd_kern.h> #include <linux/vt_kern.h>
@@ -811,18 +812,27 @@ void start_tty(struct tty_struct *tty) }EXPORT_SYMBOL (start_tty); -static void tty_update_time (struct timespec64 *time) +static void tty_update_time (struct tty_struct *tty,int tstamp) { +struct tty_file_private *priv;time64_t sec =ktime_get_real_seconds (); -/* -* We only care if the two values differ in anything other than the -* lower three bits (i.e every 8 seconds). If so, then we can update -* the time of the tty device, otherwise it could be construded as a -* security leak to let userspace know the exact timing of the tty. -*/ -if ((sec ^ time->tv_sec) & ~7 ) - time->tv_sec = sec; +spin_lock (&tty->files_lock); +list_for_each_entry (priv, &tty->tty_files, list) { +struct file *filp = priv->file; +struct inode *inode =file_inode (filp); +struct timespec64 *time = tstamp ==S_MTIME ? &inode->i_mtime : &inode->i_atime; + +/* +* We only care if the two values differ in anything other than the +* lower three bits (i.e every 8 seconds). If so, then we can update +* the time of the tty device, otherwise it could be construded as a +* security leak to let userspace know the exact timing of the tty. +*/ +if ((sec ^ time->tv_sec) & ~7 ) + time->tv_sec = sec; + } +spin_unlock (&tty->files_lock); }/*
@@ -928,7 +938,7 @@ static ssize_t tty_read(struct kiocb *iocb, struct iov_iter *to)tty_ldisc_deref (ld);if (i >0 ) -tty_update_time (&inode->i_atime); +tty_update_time (tty,S_ATIME );return i; }
@@ -1036,7 +1046,7 @@ static inline ssize_t do_tty_write(cond_resched (); }if (written) { -tty_update_time (&file_inode (file)->i_mtime); +tty_update_time (tty,S_MTIME ); ret = written; } out:
Signature · 1 line
2.39.2
PATCH 2/2 selftests: tty: add selftest for tty timestamp updates
--- tools/testing/selftests/Makefile | 1 + tools/testing/selftests/tty/.gitignore | 2 + tools/testing/selftests/tty/Makefile | 5 ++ .../testing/selftests/tty/tty_tstamp_update.c | 88 +++++++++++++++++++ 4 files changed, 96 insertions(+) create mode 100644 tools/testing/selftests/tty/.gitignore create mode 100644 tools/testing/selftests/tty/Makefile create mode 100644 tools/testing/selftests/tty/tty_tstamp_update.c
tools/testing/selftests/Makefile +1 -0
@@ -84,6 +84,7 @@ TARGETS += timers endif TARGETS += tmpfs TARGETS += tpm2 +TARGETS += tty TARGETS += user TARGETS += vDSO TARGETS += mm
tools/testing/selftests/tty/.gitignore new file +2 -0
@@ -0,0 +1,2 @@ +# SPDX-License-Identifier: GPL-2.0-only +tty_tstamp_update
tools/testing/selftests/tty/Makefile new file +5 -0
@@ -0,0 +1,5 @@ +# SPDX-License-Identifier: GPL-2.0 +CFLAGS = -O2 -Wall +TEST_GEN_PROGS := tty_tstamp_update + +include ../lib.mk
tools/testing/selftests/tty/tty_tstamp_update.c new file +88 -0
88 lines not on this page · raw message
Signature · 1 line
2.39.2
Jiri Slaby
maintainerQuoted 42 lines · Michal Sekletar
User space applications watch for timestamp changes on character device files in order to determine idle time of a given terminal session. For example, "w" program uses this information to populate the IDLE column of its output [1]. Similarly, systemd-logind has optional feature where it uses atime of the tty character device to determine if there was activity on the terminal associated with the logind's session object. If there was no activity for a configured period of time then logind will terminate such session [2].
Now, usually (e.g. bash running on the terminal) the use of the terminal will update timestamps (atime and mtime) on the corresponding terminal character device. However, if access to the terminal, e.g.
/dev/pts/0, is performed through magic character device/dev/ttythen such access obviously changes the state of the terminal, however timestamps on the device that correspond to the terminal (/dev/pts/0) are not updated.This patch makes sure that we update timestamps on *all* character devices that correspond to the given tty, because outside observers (w, systemd-logind) are maybe checking these timestamps. Obviously, they can not check timestamps on
/dev/ttyas that has per-process meaning.[1] https://gitlab.com/procps-ng/procps/-/blob/v4.0.0/w.c#L286
[2] https://github.com/systemd/systemd/blob/v252/NEWS#L477Signed-off-by Michal Sekletar <[email protected]>--- drivers/tty/tty_io.c | 32 +++++++++++++++++++++----------- 1 file changed, 21 insertions(+), 11 deletions(-)drivers/tty/tty_io.c +1
-0@@ -101,6 +101,7 @@#include <linux/compat.h> #include <linux/uaccess.h> #include <linux/termios_internal.h> +#include <linux/fs.h> #include <linux/kbd_kern.h> #include <linux/vt_kern.h>
drivers/tty/tty_io.c void start_tty(struct tty_struct *tty) · line 815
Quoted 5 lines
drivers/tty/tty_io.c +1
-1@@ -811,18 +812,27 @@ void start_tty(struct tty_struct *tty) }EXPORT_SYMBOL (start_tty); -static void tty_update_time (struct timespec64 *time) +static void tty_update_time (struct tty_struct *tty,int tstamp)
Why not enum file_time_flags then?
And "tstamp" sounds weird for what it is. It should be something like "time" or "time_flag". Or make it simply "bool mtime". And call it with true/false.
drivers/tty/tty_io.c void start_tty(struct tty_struct *tty) · line 818
Quoted 3 lines
drivers/tty/tty_io.c +1
-0{ +struct tty_file_private *priv;time64_t sec =ktime_get_real_seconds ();
Likely should be switched to have a reverse xmas tree.
drivers/tty/tty_io.c void start_tty(struct tty_struct *tty) · line 820
Quoted 10 lines
drivers/tty/tty_io.c +1
-8-/* -* We only care if the two values differ in anything other than the -* lower three bits (i.e every 8 seconds). If so, then we can update -* the time of the tty device, otherwise it could be construded as a -* security leak to let userspace know the exact timing of the tty. -*/ -if ((sec ^ time->tv_sec) & ~7 ) - time->tv_sec = sec; +spin_lock (&tty->files_lock);
Note: this should be fine wrt write lock. Have you tried running w/ lockdep enabled?
drivers/tty/tty_io.c void start_tty(struct tty_struct *tty) · line 822
Quoted 2 lines
drivers/tty/tty_io.c +2
-0+list_for_each_entry (priv, &tty->tty_files, list) { +struct file *filp = priv->file;
I think you can inline the above ^^ to the bellow vv.
drivers/tty/tty_io.c void start_tty(struct tty_struct *tty) · line 824
Quoted 2 lines
drivers/tty/tty_io.c +2
-0+struct inode *inode =file_inode (filp); +struct timespec64 *time = tstamp ==S_MTIME ? &inode->i_mtime : &inode->i_atime;
So you'd have:
struct inode *inode = file_inode(priv->file);
struct timespec64 *time = mtime ? &inode->i_mtime : &inode->i_atime;
Quoted 32 lines
drivers/tty/tty_io.c +13
-2+ +/* +* We only care if the two values differ in anything other than the +* lower three bits (i.e every 8 seconds). If so, then we can update +* the time of the tty device, otherwise it could be construded as a +* security leak to let userspace know the exact timing of the tty. +*/ +if ((sec ^ time->tv_sec) & ~7 ) + time->tv_sec = sec; + } +spin_unlock (&tty->files_lock); }/* @@ -928,7 +938,7 @@ static ssize_t tty_read(struct kiocb *iocb, struct iov_iter *to)tty_ldisc_deref (ld);if (i >0 ) -tty_update_time (&inode->i_atime); +tty_update_time (tty,S_ATIME );return i; }@@ -1036,7 +1046,7 @@ static inline ssize_t do_tty_write(cond_resched (); }if (written) { -tty_update_time (&file_inode (file)->i_mtime); +tty_update_time (tty,S_MTIME ); ret = written; } out:
Signature · 2 lines
js suse labs
PATCH v2 1/2 tty: tty_io: update timestamps on all device nodes
User space applications watch for timestamp changes on character device files in order to determine idle time of a given terminal session. For example, "w" program uses this information to populate the IDLE column of its output [1]. Similarly, systemd-logind has optional feature where it uses atime of the tty character device to determine if there was activity on the terminal associated with the logind's session object. If there was no activity for a configured period of time then logind will terminate such session [2].
Now, usually (e.g. bash running on the terminal) the use of the terminal will update timestamps (atime and mtime) on the corresponding terminal character device. However, if access to the terminal, e.g. /dev/pts/0, is performed through magic character device /dev/tty then such access obviously changes the state of the terminal, however timestamps on the device that correspond to the terminal (/dev/pts/0) are not updated.
This patch makes sure that we update timestamps on *all* character devices that correspond to the given tty, because outside observers (w, systemd-logind) are maybe checking these timestamps. Obviously, they can not check timestamps on /dev/tty as that has per-process meaning.
[1] https://gitlab.com/procps-ng/procps/-/blob/v4.0.0/w.c#L286
[2] https://github.com/systemd/systemd/blob/v252/NEWS#L477
---
v1 -> v2: Minor style tweaks based on code review comments
drivers/tty/tty_io.c | 31 ++++++++++++++++++++-----------
1 file changed, 20 insertions(+), 11 deletions(-)
drivers/tty/tty_io.c +20 -11
@@ -101,6 +101,7 @@#include <linux/compat.h> #include <linux/uaccess.h> #include <linux/termios_internal.h> +#include <linux/fs.h> #include <linux/kbd_kern.h> #include <linux/vt_kern.h>
@@ -811,18 +812,26 @@ void start_tty(struct tty_struct *tty) }EXPORT_SYMBOL (start_tty); -static void tty_update_time (struct timespec64 *time) +static void tty_update_time (struct tty_struct *tty,bool mtime) {time64_t sec =ktime_get_real_seconds (); +struct tty_file_private *priv; -/* -* We only care if the two values differ in anything other than the -* lower three bits (i.e every 8 seconds). If so, then we can update -* the time of the tty device, otherwise it could be construded as a -* security leak to let userspace know the exact timing of the tty. -*/ -if ((sec ^ time->tv_sec) & ~7 ) - time->tv_sec = sec; +spin_lock (&tty->files_lock); +list_for_each_entry (priv, &tty->tty_files, list) { +struct inode *inode =file_inode (priv->file); +struct timespec64 *time = mtime ? &inode->i_mtime : &inode->i_atime; + +/* +* We only care if the two values differ in anything other than the +* lower three bits (i.e every 8 seconds). If so, then we can update +* the time of the tty device, otherwise it could be construded as a +* security leak to let userspace know the exact timing of the tty. +*/ +if ((sec ^ time->tv_sec) & ~7 ) + time->tv_sec = sec; + } +spin_unlock (&tty->files_lock); }/*
@@ -928,7 +937,7 @@ static ssize_t tty_read(struct kiocb *iocb, struct iov_iter *to)tty_ldisc_deref (ld);if (i >0 ) -tty_update_time (&inode->i_atime); +tty_update_time (tty, false);return i; }
@@ -1036,7 +1045,7 @@ static inline ssize_t do_tty_write(cond_resched (); }if (written) { -tty_update_time (&file_inode (file)->i_mtime); +tty_update_time (tty, true); ret = written; } out:
Signature · 1 line
2.39.1
PATCH v2 2/2 selftests: tty: add selftest for tty timestamp updates
--- tools/testing/selftests/Makefile | 1 + tools/testing/selftests/tty/.gitignore | 2 + tools/testing/selftests/tty/Makefile | 5 ++ .../testing/selftests/tty/tty_tstamp_update.c | 88 +++++++++++++++++++ 4 files changed, 96 insertions(+) create mode 100644 tools/testing/selftests/tty/.gitignore create mode 100644 tools/testing/selftests/tty/Makefile create mode 100644 tools/testing/selftests/tty/tty_tstamp_update.c
tools/testing/selftests/Makefile +1 -0
@@ -86,6 +86,7 @@ TARGETS += timers endif TARGETS += tmpfs TARGETS += tpm2 +TARGETS += tty TARGETS += user TARGETS += vDSO TARGETS += mm
tools/testing/selftests/tty/.gitignore new file +2 -0
@@ -0,0 +1,2 @@ +# SPDX-License-Identifier: GPL-2.0-only +tty_tstamp_update
tools/testing/selftests/tty/Makefile new file +5 -0
@@ -0,0 +1,5 @@ +# SPDX-License-Identifier: GPL-2.0 +CFLAGS = -O2 -Wall +TEST_GEN_PROGS := tty_tstamp_update + +include ../lib.mk
tools/testing/selftests/tty/tty_tstamp_update.c new file +88 -0
88 lines not on this page · raw message
Signature · 1 line
2.39.1
Greg KH
maintainerRe: [PATCH v2 2/2] selftests: tty: add selftest for tty timestamp updates
Quoted 1 line · Michal Sekletar
Signed-off-by Michal Sekletar <[email protected]>
Sorry, but for obvious reasons, I can't take patches without any changelog text at all (and neither would you want me to.)
thanks,
greg k-h
Greg KH
maintainerQuoted 16 lines · Michal Sekletar
User space applications watch for timestamp changes on character device files in order to determine idle time of a given terminal session. For example, "w" program uses this information to populate the IDLE column of its output [1]. Similarly, systemd-logind has optional feature where it uses atime of the tty character device to determine if there was activity on the terminal associated with the logind's session object. If there was no activity for a configured period of time then logind will terminate such session [2].
Now, usually (e.g. bash running on the terminal) the use of the terminal will update timestamps (atime and mtime) on the corresponding terminal character device. However, if access to the terminal, e.g.
/dev/pts/0, is performed through magic character device/dev/ttythen such access obviously changes the state of the terminal, however timestamps on the device that correspond to the terminal (/dev/pts/0) are not updated.
Quoted 4 lines
This patch makes sure that we update timestamps on *all* character devices that correspond to the given tty, because outside observers (w, systemd-logind) are maybe checking these timestamps. Obviously, they can not check timestamps on
/dev/ttyas that has per-process meaning.
So how are you protecting this from being an information leak like we have had in the past where you could monitor how many characters were being sent to the tty through a proc file? Seems like now you can just monitor any tty node in the system and get the same information, while today you can only do it for the tty devices you have permissions for, right?
thanks,
greg k-h
A message that is not in the archive (sent off-list or not yet received): CALVzVJas7g8PrTavpQ01J4vpKtqNP7fYznMMXYEM4K5XbbXRhg@mail.gmail.com
Greg KH
maintainerQuoted 9 lines · Michal Sekletar
Quoted 6 lines · Greg KH
So how are you protecting this from being an information leak like we have had in the past where you could monitor how many characters were being sent to the tty through a proc file? Seems like now you can just monitor any tty node in the system and get the same information, while today you can only do it for the tty devices you have permissions for, right?
Quoted 7 lines
Hi Greg,
I am not protecting against it in any way, but proposed changes are only about timestamp updates which still happen in at least 8 seconds intervals so exact timing of read/writes to tty can't be inferred. Frankly, I may have misunderstood something. It would be great if you could mention a bit more details about CVE you had in mind.
Ah, I missed that this is in 8 second increments, nevermind then!
thanks,
greg k-h
Greg KH
maintainerQuoted 12 lines · Greg KH
Quoted 11 lines · Michal Sekletar
Quoted 6 lines · Greg KH
So how are you protecting this from being an information leak like we have had in the past where you could monitor how many characters were being sent to the tty through a proc file? Seems like now you can just monitor any tty node in the system and get the same information, while today you can only do it for the tty devices you have permissions for, right?
Hi Greg,
Quoted 8 lines
Quoted 5 lines
I am not protecting against it in any way, but proposed changes are only about timestamp updates which still happen in at least 8 seconds intervals so exact timing of read/writes to tty can't be inferred. Frankly, I may have misunderstood something. It would be great if you could mention a bit more details about CVE you had in mind.
Ah, I missed that this is in 8 second increments, nevermind then!
Note, I still can't take this series for the obvious reason in patch 2/2. Please fix.
thanks,
greg k-h
PATCH v3 1/2 tty: tty_io: update timestamps on all device nodes
User space applications watch for timestamp changes on character device files in order to determine idle time of a given terminal session. For example, "w" program uses this information to populate the IDLE column of its output [1]. Similarly, systemd-logind has optional feature where it uses atime of the tty character device to determine if there was activity on the terminal associated with the logind's session object. If there was no activity for a configured period of time then logind will terminate such session [2].
Now, usually (e.g. bash running on the terminal) the use of the terminal will update timestamps (atime and mtime) on the corresponding terminal character device. However, if access to the terminal, e.g. /dev/pts/0, is performed through magic character device /dev/tty then such access obviously changes the state of the terminal, however timestamps on the device that correspond to the terminal (/dev/pts/0) are not updated.
This patch makes sure that we update timestamps on *all* character devices that correspond to the given tty, because outside observers (w, systemd-logind) are maybe checking these timestamps. Obviously, they can not check timestamps on /dev/tty as that has per-process meaning.
[1] https://gitlab.com/procps-ng/procps/-/blob/v4.0.0/w.c#L286
[2] https://github.com/systemd/systemd/blob/v252/NEWS#L477
---
v2 -> v3: Added more ellaborate commit message for newly introduced kselftest
drivers/tty/tty_io.c | 31 ++++++++++++++++++++----------- 1 file changed, 20 insertions(+), 11 deletions(-)
drivers/tty/tty_io.c +20 -11
@@ -101,6 +101,7 @@#include <linux/compat.h> #include <linux/uaccess.h> #include <linux/termios_internal.h> +#include <linux/fs.h> #include <linux/kbd_kern.h> #include <linux/vt_kern.h>
@@ -811,18 +812,26 @@ void start_tty(struct tty_struct *tty) }EXPORT_SYMBOL (start_tty); -static void tty_update_time (struct timespec64 *time) +static void tty_update_time (struct tty_struct *tty,bool mtime) {time64_t sec =ktime_get_real_seconds (); +struct tty_file_private *priv; -/* -* We only care if the two values differ in anything other than the -* lower three bits (i.e every 8 seconds). If so, then we can update -* the time of the tty device, otherwise it could be construded as a -* security leak to let userspace know the exact timing of the tty. -*/ -if ((sec ^ time->tv_sec) & ~7 ) - time->tv_sec = sec; +spin_lock (&tty->files_lock); +list_for_each_entry (priv, &tty->tty_files, list) { +struct inode *inode =file_inode (priv->file); +struct timespec64 *time = mtime ? &inode->i_mtime : &inode->i_atime; + +/* +* We only care if the two values differ in anything other than the +* lower three bits (i.e every 8 seconds). If so, then we can update +* the time of the tty device, otherwise it could be construded as a +* security leak to let userspace know the exact timing of the tty. +*/ +if ((sec ^ time->tv_sec) & ~7 ) + time->tv_sec = sec; + } +spin_unlock (&tty->files_lock); }/*
@@ -928,7 +937,7 @@ static ssize_t tty_read(struct kiocb *iocb, struct iov_iter *to)tty_ldisc_deref (ld);if (i >0 ) -tty_update_time (&inode->i_atime); +tty_update_time (tty, false);return i; }
@@ -1036,7 +1045,7 @@ static inline ssize_t do_tty_write(cond_resched (); }if (written) { -tty_update_time (&file_inode (file)->i_mtime); +tty_update_time (tty, true); ret = written; } out:
Signature · 1 line
2.40.0
PATCH v3 2/2 selftests: tty: add selftest for tty timestamp updates
Add new test case which checks that timestamp updates on actual terminal character device (e.g. /dev/pts/0) happen even if the terminal is accessed via magic /dev/tty file.
--- tools/testing/selftests/Makefile | 1 + tools/testing/selftests/tty/.gitignore | 2 + tools/testing/selftests/tty/Makefile | 5 ++ .../testing/selftests/tty/tty_tstamp_update.c | 88 +++++++++++++++++++ 4 files changed, 96 insertions(+) create mode 100644 tools/testing/selftests/tty/.gitignore create mode 100644 tools/testing/selftests/tty/Makefile create mode 100644 tools/testing/selftests/tty/tty_tstamp_update.c
tools/testing/selftests/Makefile +1 -0
@@ -86,6 +86,7 @@ TARGETS += timers endif TARGETS += tmpfs TARGETS += tpm2 +TARGETS += tty TARGETS += user TARGETS += vDSO TARGETS += mm
tools/testing/selftests/tty/.gitignore new file +2 -0
@@ -0,0 +1,2 @@ +# SPDX-License-Identifier: GPL-2.0-only +tty_tstamp_update
tools/testing/selftests/tty/Makefile new file +5 -0
@@ -0,0 +1,5 @@ +# SPDX-License-Identifier: GPL-2.0 +CFLAGS = -O2 -Wall +TEST_GEN_PROGS := tty_tstamp_update + +include ../lib.mk
tools/testing/selftests/tty/tty_tstamp_update.c new file +88 -0
88 lines not on this page · raw message
Signature · 1 line
2.40.0