y2038: make CONFIG_64BIT_TIME unconditional
Subsystems (8)
PATCH 1/2 y2038: make CONFIG_64BIT_TIME unconditional
As Stepan Golosunov points out, we made a small mistake in the get_timespec64() function in the kernel. It was originally added under the assumption that CONFIG_64BIT_TIME would get enabled on all 32-bit and 64-bit architectures, but when I did the conversion, I only turned it on for 32-bit ones.
The effect is that the get_timespec64() function never clears the upper half of the tv_nsec field for 32-bit tasks in compat mode. Clearing this is required for POSIX compliant behavior of functions that pass a 'timespec' structure with a 64-bit tv_sec and a 32-bit tv_nsec, plus uninitialized padding.
The easiest fix for linux-5.1 is to just make the Kconfig symbol unconditional, as it was originally intended. As a follow-up, we should remove any #ifdef CONFIG_64BIT_TIME completely.
---
Please apply this one as a bugfix for 5.1
--- arch/Kconfig | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
arch/Kconfig +1 -1
@@ -764,7 +764,7 @@ config COMPAT_OLD_SIGACTION bool config 64BIT_TIME - def_bool ARCH_HAS_64BIT_TIME + def_bool y help This should be selected by all architectures that need to support new system calls with a 64-bit time_t. This is relevant on all 32-bit
Signature · 1 line
2.20.0
PATCH 2/2 y2038: remove CONFIG_64BIT_TIME
The CONFIG_64BIT_TIME option is defined on all architectures, and can be removed for simplicity now.
---
This would make sense for 5.2, or could be part of a later cleanup series when I have more patches for the remaining y2038 bits
--- arch/Kconfig | 8 -------- fs/aio.c | 2 +- ipc/syscall.c | 2 +- kernel/time/hrtimer.c | 2 +- kernel/time/time.c | 4 ++-- net/socket.c | 2 +- 6 files changed, 6 insertions(+), 14 deletions(-)
arch/Kconfig +0 -8
@@ -763,14 +763,6 @@ config OLD_SIGACTION config COMPAT_OLD_SIGACTION bool -config 64BIT_TIME - def_bool y - help - This should be selected by all architectures that need to support - new system calls with a 64-bit time_t. This is relevant on all 32-bit - architectures, and 64-bit architectures as part of compat syscall - handling. - config COMPAT_32BIT_TIME def_bool !64BIT || COMPAT help
fs/aio.c +1 -1
@@ -2057,7 +2057,7 @@ static long do_io_getevents(aio_context_t ctx_id,* specifies an infinite timeout. Note that the timeout pointed to by * timeout is relative. Will fail with -ENOSYS if not implemented. */ -#if !defined(CONFIG_64BIT_TIME ) || defined(CONFIG_64BIT ) +#ifdef CONFIG_64BIT SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id, long, min_nr,
ipc/syscall.c +1 -1
@@ -30,7 +30,7 @@ int ksys_ipc(unsigned int call, int first, unsigned long second,return ksys_semtimedop (first, (struct sembuf __user *)ptr, second,NULL );case SEMTIMEDOP : -if (IS_ENABLED (CONFIG_64BIT ) || !IS_ENABLED (CONFIG_64BIT_TIME )) +if (IS_ENABLED (CONFIG_64BIT ))return ksys_semtimedop (first, ptr, second, (const struct __kernel_timespec __user *)fifth);else if (IS_ENABLED (CONFIG_COMPAT_32BIT_TIME ))
kernel/time/hrtimer.c +1 -1
@@ -1749,7 +1749,7 @@ long hrtimer_nanosleep(const struct timespec64 *rqtp,return ret; } -#if !defined(CONFIG_64BIT_TIME ) || defined(CONFIG_64BIT ) +#ifdef CONFIG_64BIT SYSCALL_DEFINE2(nanosleep, struct __kernel_timespec __user *, rqtp, struct __kernel_timespec __user *, rmtp)
kernel/time/time.c +2 -2
@@ -264,7 +264,7 @@ COMPAT_SYSCALL_DEFINE2(settimeofday, struct old_timeval32 __user *, tv, }#endif -#if !defined(CONFIG_64BIT_TIME ) || defined(CONFIG_64BIT ) +#ifdef CONFIG_64BIT SYSCALL_DEFINE1(adjtimex, struct __kernel_timex __user *, txc_p) {struct __kernel_timex txc;/* Local copy of parameter */
@@ -871,7 +871,7 @@ int get_timespec64(struct timespec64 *ts, ts->tv_sec = kts.tv_sec;/* Zero out the padding for 32 bit systems or in compat mode */ -if (IS_ENABLED (CONFIG_64BIT_TIME ) &&in_compat_syscall ()) +if (in_compat_syscall ()) kts.tv_nsec &=0xFFFFFFFFUL ; ts->tv_nsec = kts.tv_nsec;
net/socket.c +1 -1
@@ -2787,7 +2787,7 @@ SYSCALL_DEFINE2(socketcall, int, call, unsigned long __user *, args) a[2], true);break ;case SYS_RECVMMSG : -if (IS_ENABLED (CONFIG_64BIT ) || !IS_ENABLED (CONFIG_64BIT_TIME )) +if (IS_ENABLED (CONFIG_64BIT )) err = __sys_recvmmsg(a0, (struct mmsghdr __user *)a1, a[2], a[3], (struct __kernel_timespec __user *)a[4],
Signature · 7 lines
2.20.0 -- To unsubscribe, send a message with 'unsubscribe linux-aio' in the body to [email protected]. For more info on Linux AIO, see: http://www.kvack.org/aio/ Don't email: <a href=mailto:"[email protected]">[email protected]</a>
Hi Arnd,
Quoted 11 lines
As Stepan Golosunov points out, we made a small mistake in the
get_timespec64()function in the kernel. It was originally added under the assumption thatCONFIG_64BIT_TIMEwould get enabled on all 32-bit and 64-bit architectures, but when I did the conversion, I only turned it on for 32-bit ones.The effect is that the
get_timespec64()function never clears the upper half of thetv_nsecfield for 32-bit tasks in compat mode. Clearing this is required for POSIX compliant behavior of functions that pass a 'timespec' structure with a 64-bittv_secand a 32-bittv_nsec, plus uninitialized padding.
On my 32 bit ARM setup (with CONFIG_COMPAT_32BIT_TIME=y) for Y2038 test:
- I do use clock_settime64 [1] to set 64 bit tv_sec with 32 bit tv_nsec and 32 bit unnamed padding.
- I also may use clock_settime32 as a fallback (but it will not work beyond Y2038) if the 64 bit version is not available for any reason.
In the get_timespec64() the in_compat_syscall() returns 0 [2], so the padding bits are not cleared.
This imposes the in glibc requirement to zero the padding before passing it to the Linux kernel.
At least for my setup (32bit ARM with 64 bit time support) this patch is not fixing anything.
Moreover, the described above problem doesn't matter on native 64 bit systems as there both fields are 64 bit (no padding required).
Note:
[1] -
https://elixir.bootlin.com/linux/v5.1-rc6/source/arch/arm/tools/syscall.tbl#L421
[2] - include/linux/compat.h -> CONFIG_COMPAT is not defined - as a result in_compat_syscall() returns 0.
Quoted 27 lines
The easiest fix for linux-5.1 is to just make the
Kconfigsymbol unconditional, as it was originally intended. As a follow-up, we should remove any #ifdefCONFIG_64BIT_TIMEcompletely.Link:
https://lore.kernel.org/lkml/[email protected]/
Cc: Lukasz Majewski <[email protected]> Cc: Stepan Golosunov <[email protected]> Signed-off-by: Arnd Bergmann <[email protected]>
---
Please apply this one as a bugfix for 5.1--- arch/Kconfig | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)arch/Kconfig +1
-1@@ -764,7 +764,7 @@ config COMPAT_OLD_SIGACTION bool config 64BIT_TIME - def_bool ARCH_HAS_64BIT_TIME + def_bool y help This should be selected by all architectures that need to
Quoted 2 lines
support new system calls with a 64-bit
time_t. This is relevant on all 32-bit
Best regards,
Lukasz Majewski
--
DENX Software Engineering GmbH, Managing Director: Wolfgang Denk
HRB 165235 Munich, Office: Kirchenstr.5, D-82194 Groebenzell, Germany
Phone: (+49)-8142-66989-59 Fax: (+49)-8142-66989-80 Email: [email protected]
27.04.2019 в 00:46:53 +0200 Lukasz Majewski написал:
Quoted 8 lines
Hi Arnd,
Quoted 6 lines
As Stepan Golosunov points out, we made a small mistake in the
get_timespec64()function in the kernel. It was originally added under the assumption thatCONFIG_64BIT_TIMEwould get enabled on all 32-bit and 64-bit architectures, but when I did the conversion, I only turned it on for 32-bit ones.
Quoted 8 lines
Quoted 5 lines
The effect is that the
get_timespec64()function never clears the upper half of thetv_nsecfield for 32-bit tasks in compat mode. Clearing this is required for POSIX compliant behavior of functions that pass a 'timespec' structure with a 64-bittv_secand a 32-bittv_nsec, plus uninitialized padding.At least for my setup (32bit ARM with 64 bit time support) this patch is not fixing anything.
The patch is not supposed to fix anything on 32-bit architectures as in-kernel struct timespec64 has 32-bit tv_nsec there. Thus truncation should happen automatically. I also missed that fact when I was reading get_timespec64 code.
(I am wondering whether such trucation is undefined behaviour in C and whether there should be sign-extension instead of zeroing-out for the in_compat_syscall() case though.)
Quoted 28 lines
Quoted 28 lines
The easiest fix for linux-5.1 is to just make the
Kconfigsymbol unconditional, as it was originally intended. As a follow-up, we should remove any #ifdefCONFIG_64BIT_TIMEcompletely.Link:
https://lore.kernel.org/lkml/[email protected]/
Cc: Lukasz Majewski <[email protected]> Cc: Stepan Golosunov <[email protected]> Signed-off-by: Arnd Bergmann <[email protected]>
---
Please apply this one as a bugfix for 5.1--- arch/Kconfig | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)arch/Kconfig +1
-1@@ -764,7 +764,7 @@ config COMPAT_OLD_SIGACTION bool config 64BIT_TIME - def_bool ARCH_HAS_64BIT_TIME + def_bool y help This should be selected by all architectures that need tosupport new system calls with a 64-bit
time_t. This is relevant on all 32-bit
Hi Stepan,
Quoted 15 lines
27.04.2019 в 00:46:53 +0200 Lukasz Majewski написал:
Quoted 13 lines
Hi Arnd,
Quoted 11 lines
As Stepan Golosunov points out, we made a small mistake in the
get_timespec64()function in the kernel. It was originally added under the assumption thatCONFIG_64BIT_TIMEwould get enabled on all 32-bit and 64-bit architectures, but when I did the conversion, I only turned it on for 32-bit ones.The effect is that the
get_timespec64()function never clears the upper half of thetv_nsecfield for 32-bit tasks in compat mode. Clearing this is required for POSIX compliant behavior of functions that pass a 'timespec' structure with a 64-bittv_secand a 32-bittv_nsec, plus uninitialized padding.
Quoted 7 lines
Quoted 2 lines
At least for my setup (32bit ARM with 64 bit time support) this patch is not fixing anything.
The patch is not supposed to fix anything on 32-bit architectures as in-kernel struct timespec64 has 32-bit
tv_nsecthere. Thus truncation should happen automatically. I also missed that fact when I was readingget_timespec64code.
Yes. You are right. The tv_nsec is 32 bit.
Quoted 2 lines
(I am wondering whether such trucation is undefined behaviour in C
According to [1] - Chapter 6.3.1.3 - Point 3 it is implementation-defined.
Quoted 3 lines
and
whether there should be sign-extension instead of zeroing-out for thein_compat_syscall()case though.)
What I've found is that "typically" the high order bits are discarded.
However, it is still "implementation dependent".
Quoted 22 lines
Quoted 22 lines
Quoted 22 lines
The easiest fix for linux-5.1 is to just make the
Kconfigsymbol unconditional, as it was originally intended. As a follow-up, we should remove any #ifdefCONFIG_64BIT_TIMEcompletely.Link:
https://lore.kernel.org/lkml/[email protected]/
Cc: Lukasz Majewski <[email protected]> Cc: Stepan Golosunov <[email protected]> Signed-off-by: Arnd Bergmann <[email protected]> ---
Please apply this one as a bugfix for 5.1--- arch/Kconfig | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)arch/Kconfig +0
-0@@ -764,7 +764,7 @@ config COMPAT_OLD_SIGACTION bool
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
arch/Kconfig +1
-1config 64BIT_TIME - def_bool ARCH_HAS_64BIT_TIME + def_bool y help This should be selected by all architectures that needto support new system calls with a 64-bit
time_t. This is relevant on all 32-bit
Note:
[1] - http://www.open-std.org/jtc1/sc22/wg14/www/docs/n1570.pdf
Best regards,
Lukasz Majewski
--
DENX Software Engineering GmbH, Managing Director: Wolfgang Denk
HRB 165235 Munich, Office: Kirchenstr.5, D-82194 Groebenzell, Germany
Phone: (+49)-8142-66989-59 Fax: (+49)-8142-66989-80 Email: [email protected]
Quoted 5 lines · Lukasz Majewski
Quoted 2 lines
27.04.2019 в 00:46:53 +0200 Lukasz Majewski написал:
(I am wondering whether such trucation is undefined behaviour in CAccording to [1] - Chapter 6.3.1.3 - Point 3 it is implementation-defined.
The kernel relies on the sane behavior for integer overflow in many places already, and it is built with -fno-strict-overflow to also make sure gcc doesn't optimize cases that would be undefined otherwise.
Quoted 7 lines
Quoted 3 lines
and
whether there should be sign-extension instead of zeroing-out for thein_compat_syscall()case though.)What I've found is that "typically" the high order bits are discarded.
However, it is still "implementation dependent".
I think the question was whether we should use
kts.tv_nsec = (int)kts.tv_nsec;
instead of
kts.tv_nsec &= 0xfffffffful;
Both have a clearly defined meaning in the C dialect we use in the kernel, but differ in the upper 32 bits for negative input values.
I would say that using sign-extension indeed makes more sense here, but we don't need to change it for linux-5.1, since none of the callers of get_timespec64() care -- any negative 32-bit tv_nsec value results in -EINVAL, including the utimensat() syscall that has two special cases outside of the 0...999999999 range.
Arnd
Quoted 16 lines · Arnd Bergmann
As Stepan Golosunov points out, we made a small mistake in the
get_timespec64()function in the kernel. It was originally added under the assumption thatCONFIG_64BIT_TIMEwould get enabled on all 32-bit and 64-bit architectures, but when I did the conversion, I only turned it on for 32-bit ones.The effect is that the
get_timespec64()function never clears the upper half of thetv_nsecfield for 32-bit tasks in compat mode. Clearing this is required for POSIX compliant behavior of functions that pass a 'timespec' structure with a 64-bittv_secand a 32-bittv_nsec, plus uninitialized padding.The easiest fix for linux-5.1 is to just make the
Kconfigsymbol unconditional, as it was originally intended. As a follow-up, we should remove any #ifdefCONFIG_64BIT_TIMEcompletely.
Quoted 6 lines
Cc Lukasz Majewski <[email protected]>Cc Stepan Golosunov <[email protected]>Signed-off-by Arnd Bergmann <[email protected]>---
Please apply this one as a bugfix for 5.1
Can you provide a 'Fixes: ....' tag please?
Thanks,
tglx
Quoted 19 lines · Thomas Gleixner
Quoted 16 lines · Arnd Bergmann
As Stepan Golosunov points out, we made a small mistake in the
get_timespec64()function in the kernel. It was originally added under the assumption thatCONFIG_64BIT_TIMEwould get enabled on all 32-bit and 64-bit architectures, but when I did the conversion, I only turned it on for 32-bit ones.The effect is that the
get_timespec64()function never clears the upper half of thetv_nsecfield for 32-bit tasks in compat mode. Clearing this is required for POSIX compliant behavior of functions that pass a 'timespec' structure with a 64-bittv_secand a 32-bittv_nsec, plus uninitialized padding.The easiest fix for linux-5.1 is to just make the
Kconfigsymbol unconditional, as it was originally intended. As a follow-up, we should remove any #ifdefCONFIG_64BIT_TIMEcompletely.
Quoted 8 lines
Quoted 6 lines
Cc Lukasz Majewski <[email protected]>Cc Stepan Golosunov <[email protected]>Signed-off-by Arnd Bergmann <[email protected]>---
Please apply this one as a bugfix for 5.1Can you provide a 'Fixes: ....' tag please?
Ok, resent both patches now. I also took the chance to add a clarification for the point that Lukasz missed on the first submission.
Arnd