Skip to contentSkip to conversation

tty: n_gsm: fix use-after-free in gsmtty_install

RFC New Kito Xu (veritas501) 2 participants 3 messages Apr 15 – Apr 20 TTY layer and serial drivers

lore
Share & tools
lore
#0 lore raw

RFC 1/1 tty: n_gsm: fix use-after-free in gsmtty_install

gsmtty_install() reads gsm_mux[] without holding gsm_mux_lock and defers mux_get() about 40 lines later. A concurrent gsmld_close() can free the mux through gsm_cleanup_mux() -> mux_put() -> kfree(gsm) in that window, leading to a use-after-free when gsmtty_install() later dereferences the stale pointer.

race condition:

         cpu 0                          |            cpu 1
    gsmtty_install()                    |   gsmld_close()
    gsm = gsm_mux[mux] // no lock      |
    // CFS preempts here                |   gsm_cleanup_mux(gsm)
                                        |   gsm->dead = true
                                        |   mux_put(gsm)
                                        |    -> kfree(gsm)
    gsm->dead           // UAF!        |
    mutex_lock(&gsm->mutex) // UAF!    |

Acquire gsm_mux_lock before reading gsm_mux[] and take the refcount via kref_get() while still holding the lock, so the mux cannot be freed between lookup and refcount increment.

Fixes 86176ed90545 ("TTY: n_gsm, use tty_port_install")
---
 drivers/tty/n_gsm.c | 23 ++++++++++++++++-------
 1 file changed, 16 insertions(+), 7 deletions(-)
drivers/tty/n_gsm.c +16 -7
index c13e050de83b..6519f1c92fc5 100644
@@ -4288,14 +4288,20 @@ static int gsmtty_install(struct tty_driver *driver, struct tty_struct *tty)
 
 	if (mux >= MAX_MUX)
 		return -ENXIO;
-	/* FIXME: we need to lock gsm_mux for lifetimes of ttys eventually */
-	if (gsm_mux[mux] == NULL)
-		return -EUNATCH;
 	if (line == 0 || line > 61)	/* 62/63 reserved */
 		return -ECHRNG;
+
+	/* Acquire gsm_mux_lock to prevent concurrent gsmld_close() from
+	 * freeing the mux between reading gsm_mux[] and taking a refcount.
+	 */
+	spin_lock(&gsm_mux_lock);
 	gsm = gsm_mux[mux];
-	if (gsm->dead)
-		return -EL2HLT;
+	if (!gsm || gsm->dead) {
+		spin_unlock(&gsm_mux_lock);
+		return gsm ? -EL2HLT : -EUNATCH;
+	}
+	kref_get(&gsm->ref);
+	spin_unlock(&gsm_mux_lock);
 	/* If DLCI 0 is not yet fully open return an error.
 	This is ok from a locking
 	perspective as we don't have to worry about this
@@ -4309,8 +4315,10 @@ static int gsmtty_install(struct tty_driver *driver, struct tty_struct *tty)
 		if (dlci0->state == DLCI_OPENING)
 			wait_event(gsm->event, dlci0->state != DLCI_OPENING);
 
-		if (dlci0->state != DLCI_OPEN)
+		if (dlci0->state != DLCI_OPEN) {
+			mux_put(gsm);
 			return -EL2NSYNC;
+		}
 
 		mutex_lock(&gsm->mutex);
 	}
@@ -4322,6 +4330,7 @@ static int gsmtty_install(struct tty_driver *driver, struct tty_struct *tty)
 	}
 	if (dlci == NULL) {
 		mutex_unlock(&gsm->mutex);
+		mux_put(gsm);
 		return -ENOMEM;
 	}
 	ret = tty_port_install(&dlci->port, driver, tty);
@@ -4329,12 +4338,12 @@ static int gsmtty_install(struct tty_driver *driver, struct tty_struct *tty)
 		if (alloc)
 			dlci_put(dlci);
 		mutex_unlock(&gsm->mutex);
+		mux_put(gsm);
 		return ret;
 	}
 
 	dlci_get(dlci);
 	dlci_get(gsm->dlci[0]);
-	mux_get(gsm);
 	tty->driver_data = dlci;
 	mutex_unlock(&gsm->mutex);
 
Signature · 1 line
2.43.0
#1 lore raw

PATCH 1/1 tty: n_gsm: fix use-after-free in gsmtty_install

gsmtty_install() reads gsm_mux[] without holding gsm_mux_lock and defers mux_get() about 40 lines later. A concurrent gsmld_close() can free the mux through gsm_cleanup_mux() -> mux_put() -> kfree(gsm) in that window, leading to a use-after-free when gsmtty_install() later dereferences the stale pointer.

race condition:

         cpu 0                          |            cpu 1
    gsmtty_install()                    |   gsmld_close()
    gsm = gsm_mux[mux] // no lock      |
    // CFS preempts here                |   gsm_cleanup_mux(gsm)
                                        |   gsm->dead = true
                                        |   mux_put(gsm)
                                        |    -> kfree(gsm)
    gsm->dead           // UAF!        |
    mutex_lock(&gsm->mutex) // UAF!    |

KASAN report:

 BUG: KASAN: slab-use-after-free in gsmtty_install+0x6cf/0x830
 Read of size 1 at addr ffff88800fd440ac by task poc/170

 CPU: 0 UID: 0 PID: 170 Comm: poc Not tainted 7.0.0-rc7-next-20260410+ #20
 Call Trace:
  <TASK>
  dump_stack_lvl+0x64/0x80
  print_report+0xd0/0x5e0
  kasan_report+0xce/0x100
  gsmtty_install+0x6cf/0x830
  tty_init_dev.part.0+0x92/0x4a0
  tty_open+0x8ab/0x1050
  chrdev_open+0x1ec/0x5e0
  do_dentry_open+0x419/0x1260
  vfs_open+0x79/0x350
  path_openat+0x212c/0x3a70
  do_file_open+0x1d2/0x400
  do_sys_openat2+0xdc/0x170
  __x64_sys_openat+0x122/0x1e0
  do_syscall_64+0x64/0x680
  entry_SYSCALL_64_after_hwframe+0x76/0x7e
  </TASK>

 Allocated by task 169 on cpu 0 at 3.824684s:
 Freed by task 169 on cpu 0 at 3.892012s:

 The buggy address belongs to the object at ffff88800fd44000
  which belongs to the cache kmalloc-1k of size 1024
 The buggy address is located 172 bytes inside of
  freed 1024-byte region [ffff88800fd44000, ffff88800fd44400)

Acquire gsm_mux_lock before reading gsm_mux[] and take the refcount via kref_get() while still holding the lock, so the mux cannot be freed between lookup and refcount increment.

Fixes 86176ed90545 ("TTY: n_gsm, use tty_port_install")
---
 drivers/tty/n_gsm.c | 23 ++++++++++++++++-------
 1 file changed, 16 insertions(+), 7 deletions(-)
drivers/tty/n_gsm.c +16 -7
index c13e050de83b..6519f1c92fc5 100644
@@ -4288,14 +4288,20 @@ static int gsmtty_install(struct tty_driver *driver, struct tty_struct *tty)
 
 	if (mux >= MAX_MUX)
 		return -ENXIO;
-	/* FIXME: we need to lock gsm_mux for lifetimes of ttys eventually */
-	if (gsm_mux[mux] == NULL)
-		return -EUNATCH;
 	if (line == 0 || line > 61)	/* 62/63 reserved */
 		return -ECHRNG;
+
+	/* Acquire gsm_mux_lock to prevent concurrent gsmld_close() from
+	 * freeing the mux between reading gsm_mux[] and taking a refcount.
+	 */
+	spin_lock(&gsm_mux_lock);
 	gsm = gsm_mux[mux];
-	if (gsm->dead)
-		return -EL2HLT;
+	if (!gsm || gsm->dead) {
+		spin_unlock(&gsm_mux_lock);
+		return gsm ? -EL2HLT : -EUNATCH;
+	}
+	kref_get(&gsm->ref);
+	spin_unlock(&gsm_mux_lock);
 	/* If DLCI 0 is not yet fully open return an error.
 	This is ok from a locking
 	perspective as we don't have to worry about this
@@ -4309,8 +4315,10 @@ static int gsmtty_install(struct tty_driver *driver, struct tty_struct *tty)
 		if (dlci0->state == DLCI_OPENING)
 			wait_event(gsm->event, dlci0->state != DLCI_OPENING);
 
-		if (dlci0->state != DLCI_OPEN)
+		if (dlci0->state != DLCI_OPEN) {
+			mux_put(gsm);
 			return -EL2NSYNC;
+		}
 
 		mutex_lock(&gsm->mutex);
 	}
@@ -4322,6 +4330,7 @@ static int gsmtty_install(struct tty_driver *driver, struct tty_struct *tty)
 	}
 	if (dlci == NULL) {
 		mutex_unlock(&gsm->mutex);
+		mux_put(gsm);
 		return -ENOMEM;
 	}
 	ret = tty_port_install(&dlci->port, driver, tty);
@@ -4329,12 +4338,12 @@ static int gsmtty_install(struct tty_driver *driver, struct tty_struct *tty)
 		if (alloc)
 			dlci_put(dlci);
 		mutex_unlock(&gsm->mutex);
+		mux_put(gsm);
 		return ret;
 	}
 
 	dlci_get(dlci);
 	dlci_get(gsm->dlci[0]);
-	mux_get(gsm);
 	tty->driver_data = dlci;
 	mutex_unlock(&gsm->mutex);
 
Signature · 1 line
2.43.0
maintainer
#2 lore raw
Quoted 18 lines · On Wed, Apr 15, 2026 at 11:17:31AM +0800, Kito Xu (veritas501) wrote:

gsmtty_install() reads gsm_mux[] without holding gsm_mux_lock and defers mux_get() about 40 lines later. A concurrent gsmld_close() can free the mux through gsm_cleanup_mux() -> mux_put() -> kfree(gsm) in that window, leading to a use-after-free when gsmtty_install() later dereferences the stale pointer.

race condition:

         cpu 0                          |            cpu 1
    gsmtty_install()                    |   gsmld_close()
    gsm = gsm_mux[mux] // no lock      |
    // CFS preempts here                |   gsm_cleanup_mux(gsm)
                                        |   gsm->dead = true
                                        |   mux_put(gsm)
                                        |    -> kfree(gsm)
    gsm->dead           // UAF!        |
    mutex_lock(&gsm->mutex) // UAF!    |
Quoted 4 lines

KASAN report:

 BUG: KASAN: slab-use-after-free in gsmtty_install+0x6cf/0x830
 Read of size 1 at addr ffff88800fd440ac by task poc/170

Do you have this hardware to test with, or was this done in a virtual machine?

Quoted 29 lines
 CPU: 0 UID: 0 PID: 170 Comm: poc Not tainted 7.0.0-rc7-next-20260410+ #20
 Call Trace:
  <TASK>
  dump_stack_lvl+0x64/0x80
  print_report+0xd0/0x5e0
  kasan_report+0xce/0x100
  gsmtty_install+0x6cf/0x830
  tty_init_dev.part.0+0x92/0x4a0
  tty_open+0x8ab/0x1050
  chrdev_open+0x1ec/0x5e0
  do_dentry_open+0x419/0x1260
  vfs_open+0x79/0x350
  path_openat+0x212c/0x3a70
  do_file_open+0x1d2/0x400
  do_sys_openat2+0xdc/0x170
  __x64_sys_openat+0x122/0x1e0
  do_syscall_64+0x64/0x680
  entry_SYSCALL_64_after_hwframe+0x76/0x7e
  </TASK>

 Allocated by task 169 on cpu 0 at 3.824684s:
 Freed by task 169 on cpu 0 at 3.892012s:

 The buggy address belongs to the object at ffff88800fd44000
  which belongs to the cache kmalloc-1k of size 1024
 The buggy address is located 172 bytes inside of
  freed 1024-byte region [ffff88800fd44000, ffff88800fd44400)
Quoted 6 lines

Acquire gsm_mux_lock before reading gsm_mux[] and take the refcount via kref_get() while still holding the lock, so the mux cannot be freed between lookup and refcount increment.

Fixes 86176ed90545 ("TTY: n_gsm, use tty_port_install")

Did you test this change out?

Why was one version sent as "RFC" folllowed by this?

Quoted 12 lines
---
 drivers/tty/n_gsm.c | 23 ++++++++++++++++-------
 1 file changed, 16 insertions(+), 7 deletions(-)
drivers/tty/n_gsm.c +0 -0
index c13e050de83b..6519f1c92fc5 100644
@@ -4288,14 +4288,20 @@ static int gsmtty_install(struct tty_driver *driver, struct tty_struct *tty)
 
 	if (mux >= MAX_MUX)
 		return -ENXIO;

drivers/tty/n_gsm.c static int gsmtty_install(struct tty_driver *driver, struct tty_struct *tty) · line 4297

Quoted 10 lines
drivers/tty/n_gsm.c +5 -3
-	/* FIXME: we need to lock gsm_mux for lifetimes of ttys eventually */
-	if (gsm_mux[mux] == NULL)
-		return -EUNATCH;
 	if (line == 0 || line > 61)	/* 62/63 reserved */
 		return -ECHRNG;
+
+	/* Acquire gsm_mux_lock to prevent concurrent gsmld_close() from
+	 * freeing the mux between reading gsm_mux[] and taking a refcount.
+	 */
+	spin_lock(&gsm_mux_lock);

Wrong comment style :(

Also, how was this tested, last time this was attempted, lockups happened, hence the FIXME line.

And why isn't this the irqsave version of the lock? Why is that not needed here? Have you searched the email archives for when this was attempted in the past? What has now changed?

drivers/tty/n_gsm.c static int gsmtty_install(struct tty_driver *driver, struct tty_struct *tty) · line 4301

Quoted 6 lines
drivers/tty/n_gsm.c +3 -2
 	gsm = gsm_mux[mux];
-	if (gsm->dead)
-		return -EL2HLT;
+	if (!gsm || gsm->dead) {
+		spin_unlock(&gsm_mux_lock);
+		return gsm ? -EL2HLT : -EUNATCH;

Please spell out inline if statements like this.

drivers/tty/n_gsm.c static int gsmtty_install(struct tty_driver *driver, struct tty_struct *tty) · line 4303

Quoted 2 lines
drivers/tty/n_gsm.c +2 -0
+	}
+	kref_get(&gsm->ref);

Shouldn't this be mux_get() somehow? Open-coding it like this is confusing.

thanks,

greg k-h