summaryrefslogtreecommitdiff
path: root/include/linux/timekeeping.h
diff options
context:
space:
mode:
authorArnd Bergmann <arnd@arndb.de>2017-10-12 17:06:11 +0300
committerJason Wessel <jason.wessel@windriver.com>2018-01-25 17:40:18 +0300
commit6909e29fdefbb7aa643021279daef6ed10c81528 (patch)
tree7a84f7a177b867e0ba3f1c14dce298cd491fb3fc /include/linux/timekeeping.h
parent0296c248b440fe0ae3f08c5c2fcded795c59353e (diff)
downloadlinux-6909e29fdefbb7aa643021279daef6ed10c81528.tar.xz
kdb: use __ktime_get_real_seconds instead of __current_kernel_time
kdb is the only user of the __current_kernel_time() interface, which is not y2038 safe and should be removed at some point. The kdb code also goes to great lengths to print the time in a human-readable format from 'struct timespec', again using a non-y2038-safe re-implementation of the generic time_to_tm() code. Using __current_kernel_time() here is necessary since the regular accessors that require a sequence lock might hang when called during the xtime update. However, this is safe in the particular case since kdb is only interested in the tv_sec field that is updated atomically. In order to make this y2038-safe, I'm converting the code to the generic time64_to_tm helper, but that introduces the problem that we have no interface like __current_kernel_time() that provides a 64-bit timestamp in a lockless, safe and architecture-independent way. I have multiple ideas for how to solve that: - __ktime_get_real_seconds() is lockless, but can return incorrect results on 32-bit architectures in the special case that we are in the process of changing the time across the epoch, either during the timer tick that overflows the seconds in 2038, or while calling settimeofday. - ktime_get_real_fast_ns() would work in this context, but does require a call into the clocksource driver to return a high-resolution timestamp. This may have undesired side-effects in the debugger, since we want to limit the interactions with the rest of the kernel. - Adding a ktime_get_real_fast_seconds() based on tk_fast_mono plus tkr->base_real without the tk_clock_read() delta. Not sure about the value of adding yet another interface here. - Changing the existing ktime_get_real_seconds() to use tk_fast_mono on 32-bit architectures rather than xtime_sec. I think this could work, but am not entirely sure if this is an improvement. I picked the first of those for simplicity here. It's technically not correct but probably good enough as the time is only used for the debugging output and the race will likely never be hit in practice. Another downside is having to move the declaration into a public header file. Let me know if anyone has a different preference. Cc: Andy Shevchenko <andriy.shevchenko@linux.intel.com> Link: https://patchwork.kernel.org/patch/9775309/ Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Jason Wessel <jason.wessel@windriver.com>
Diffstat (limited to 'include/linux/timekeeping.h')
-rw-r--r--include/linux/timekeeping.h1
1 files changed, 1 insertions, 0 deletions
diff --git a/include/linux/timekeeping.h b/include/linux/timekeeping.h
index b17bcce58bc4..588a0e4b1ab9 100644
--- a/include/linux/timekeeping.h
+++ b/include/linux/timekeeping.h
@@ -31,6 +31,7 @@ struct timespec64 get_monotonic_coarse64(void);
extern void getrawmonotonic64(struct timespec64 *ts);
extern void ktime_get_ts64(struct timespec64 *ts);
extern time64_t ktime_get_seconds(void);
+extern time64_t __ktime_get_real_seconds(void);
extern time64_t ktime_get_real_seconds(void);
extern int __getnstimeofday64(struct timespec64 *tv);