firewire: core: propagate cycle-timer read failure in legacy ioctl
PATCH 1/1 firewire: core: propagate cycle-timer read failure in legacy ioctl
The legacy FW_CDEV_IOC_GET_CYCLE_TIMER handler ioctl_get_cycle_timer() reads into a local struct fw_cdev_get_cycle_timer2 via ioctl_get_cycle_timer2() and then copies tv_sec, tv_nsec and cycle_timer into the user-visible output, but it ignores the helper's return value. Since commit bacf921c42bb ("firewire: core: use guard macro to disable local IRQ") the helper returns early on fw_card_read_cycle_time() failure (for example -ENODEV after the dummy card driver is swapped in during controller removal) without writing those fields, so the wrapper copies out the uninitialized local: 12 of the 16 user-visible bytes are uninitialized kernel stack. The non-legacy ioctl already returns the error on this path.
Propagate the helper return value so no output is copied on failure.
Reproduced under KMSAN; the legacy wrapper no longer returns uninitialized cycle-timer bytes after this change.
---
Reproduction (v7.1-rc4, x86_64 QEMU, CONFIG_KMSAN=y, stack auto-init off):
With fw_card_read_cycle_time() forced to fail, the stock legacy handler returns 0 and KMSAN reports the uninitialized local being stored to the output:
BUG: KMSAN: uninit-value in ioctl_get_cycle_timer Local variable ct2 created at: ioctl_get_cycle_timer
The leaked output is local_time (8 bytes, bytes 0-7, computed from the uninitialized tv_sec/tv_nsec) and cycle_timer (4 bytes, bytes 8-11, copied verbatim). After this change the handler returns -ENODEV and the KMSAN reports are gone; a success-path control stays clean.
drivers/firewire/core-cdev.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-)
drivers/firewire/core-cdev.c +4 -1
@@ -1283,9 +1283,12 @@ static int ioctl_get_cycle_timer(struct client *client, union ioctl_arg *arg) {struct fw_cdev_get_cycle_timer *a = &arg->get_cycle_timer;struct fw_cdev_get_cycle_timer2 ct2; +int ret; ct2.clk_id =CLOCK_REALTIME ; -ioctl_get_cycle_timer2 (client, (union ioctl_arg *)&ct2); + ret =ioctl_get_cycle_timer2 (client, (union ioctl_arg *)&ct2); +if (ret <0 ) +return ret; a->local_time = ct2.tv_sec *USEC_PER_SEC + ct2.tv_nsec /NSEC_PER_USEC ; a->cycle_timer = ct2.cycle_timer;
base-commit: e24a98d6884a6e4203a77a94f070a59fcab95208
Signature · 1 line
2.53.0
Takashi Sakamoto
maintainerHi,
Thanks for the patch, however I can see some subtle gotchas.
Quoted 31 lines · Michael Bommarito
The legacy
FW_CDEV_IOC_GET_CYCLE_TIMERhandlerioctl_get_cycle_timer()reads into a local structfw_cdev_get_cycle_timer2viaioctl_get_cycle_timer2()and then copiestv_sec,tv_nsecandcycle_timerinto the user-visible output, but it ignores the helper's return value. Since commitbacf921c42bb("firewire: core: use guard macro to disable local IRQ") the helper returns early onfw_card_read_cycle_time()failure (for example-ENODEVafter the dummy card driver is swapped in during controller removal) without writing those fields, so the wrapper copies out the uninitialized local: 12 of the 16 user-visible bytes are uninitialized kernel stack. The non-legacy ioctl already returns the error on this path.Propagate the helper return value so no output is copied on failure.
Reproduced under KMSAN; the legacy wrapper no longer returns uninitialized cycle-timer bytes after this change.
Fixes bacf921c42bb ("firewire: core: use guard macro to disable local IRQ")Cc [email protected] # v6.12+Assisted-by Claude:claude-opus-4-8Signed-off-by Michael Bommarito <[email protected]>---
Reproduction (v7.1-rc4,x86_64QEMU,CONFIG_KMSAN=y, stack auto-init off):With
fw_card_read_cycle_time()forced to fail, the stock legacy handler returns 0 and KMSAN reports the uninitialized local being stored to the output:BUG: KMSAN: uninit-value in ioctl_get_cycle_timer Local variable ct2 created at: ioctl_get_cycle_timer
Quoted 7 lines
The leaked output is
local_time(8 bytes, bytes 0-7, computed from the uninitialized tv_sec/tv_nsec) andcycle_timer(4 bytes, bytes 8-11, copied verbatim). After this change the handler returns-ENODEVand the KMSAN reports are gone; a success-path control stays clean.drivers/firewire/core-cdev.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-)
The issued function, ioctl_get_cycle_timer() was introduced by a commit abfe5a01ef1e ("firewire: cdev: add more flexible cycle timer ioctl")[1]. We can see in this time the function has already ignored the return value of get_cycle_timer2(). The issue is independent of the later commit bacf921c42bb ("firewire: core: use guard macro to disable local IRQ").
Besides, userspace applications can unbind 1394 OHCI driver in any time, so it is possible to use the vulnerability reading small part of kernel stack. (Understandably, if the attacker is successful to install 1394 OHCI device to the target hardware, he or she can accomplish more jobs regardless of the issue.)
Anyway, the leak of kernel stack to userspace is not preferable itself. I would like to see your revised patch, writing on your own.
[1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=abfe5a01ef1e4
Thanks
Takashi Sakamoto
PATCH v2 1/1 firewire: core: propagate cycle-timer read failure in legacy ioctl
The legacy FW_CDEV_IOC_GET_CYCLE_TIMER handler ioctl_get_cycle_timer() reads into a local struct fw_cdev_get_cycle_timer2 via ioctl_get_cycle_timer2() and copies tv_sec, tv_nsec and cycle_timer into the user-visible output, but it has ignored the helper's return value since the ioctl was introduced. When ioctl_get_cycle_timer2() returns early on an fw_card_read_cycle_time() failure (for example -ENODEV after the dummy card driver is swapped in during controller removal) it does not populate the output, so the wrapper copies out the uninitialized
local: 12 of the 16 user-visible bytes are uninitialized kernel stack. The non-legacy ioctl already returns the error on this path. A local user can reach this by unbinding the 1394 OHCI driver while holding the character device open.
Propagate the helper return value so no output is copied on failure.
The present uninitialized read is reachable since v6.12: commit bacf921c42bb ("firewire: core: use guard macro to disable local IRQ") changed the helper's error path to return early and skip the field writes, leaving the wrapper's local uninitialized. The stable backport is therefore limited to v6.12 and later even though the ignored return value dates back to the original ioctl.
Reproduced under KMSAN; the legacy wrapper no longer returns uninitialized cycle-timer bytes after this change.
---
v2: correct the Fixes attribution per Takashi Sakamoto's review
(https://lore.kernel.org/all/[email protected]/).
ioctl_get_cycle_timer() has ignored ioctl_get_cycle_timer2()'s return value since it was added in abfe5a01ef1e, so the Fixes tag now points there rather than at bacf921c42bb. Reworded the changelog accordingly and added Sakamoto's reachability note (unbinding the 1394 OHCI driver at runtime). The stable range stays v6.12+ because the helper's error path only began returning early and skipping the field writes in bacf921c42bb; earlier kernels wrote the fields on that path. No code change from v1.
v1: https://lore.kernel.org/all/[email protected]/
Reproduction (v7.1-rc4, x86_64 QEMU, CONFIG_KMSAN=y, stack auto-init off):
With fw_card_read_cycle_time() forced to fail, the stock legacy handler returns 0 and KMSAN reports the uninitialized local being stored to the output:
BUG: KMSAN: uninit-value in ioctl_get_cycle_timer Local variable ct2 created at: ioctl_get_cycle_timer
The leaked output is local_time (8 bytes, bytes 0-7, computed from the uninitialized tv_sec/tv_nsec) and cycle_timer (4 bytes, bytes 8-11, copied verbatim). After this change the handler returns -ENODEV and the KMSAN reports are gone; a success-path control stays clean.
drivers/firewire/core-cdev.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-)
drivers/firewire/core-cdev.c +4 -1
@@ -1283,9 +1283,12 @@ static int ioctl_get_cycle_timer(struct client *client, union ioctl_arg *arg) {struct fw_cdev_get_cycle_timer *a = &arg->get_cycle_timer;struct fw_cdev_get_cycle_timer2 ct2; +int ret; ct2.clk_id =CLOCK_REALTIME ; -ioctl_get_cycle_timer2 (client, (union ioctl_arg *)&ct2); + ret =ioctl_get_cycle_timer2 (client, (union ioctl_arg *)&ct2); +if (ret <0 ) +return ret; a->local_time = ct2.tv_sec *USEC_PER_SEC + ct2.tv_nsec /NSEC_PER_USEC ; a->cycle_timer = ct2.cycle_timer;
base-commit: e24a98d6884a6e4203a77a94f070a59fcab95208
Signature · 1 line
2.53.0