]> git.baikalelectronics.ru Git - kernel.git/commitdiff
ia64: fix clock_getres(CLOCK_MONOTONIC) to report ITC frequency
authorSergei Trofimovich <slyich@gmail.com>
Sat, 20 Aug 2022 18:18:13 +0000 (19:18 +0100)
committerAndrew Morton <akpm@linux-foundation.org>
Mon, 12 Sep 2022 04:55:07 +0000 (21:55 -0700)
clock_gettime(CLOCK_MONOTONIC, &tp) is very precise on ia64 as it uses ITC
(similar to rdtsc on x86).  It's not quite a hrtimer as it is a few times
slower than 1ns.  Usually 2-3ns.

clock_getres(CLOCK_MONOTONIC, &res) never reflected that fact and reported
0.04s precision (1/HZ value).

In https://bugs.gentoo.org/596382 gstreamer's test suite failed loudly
when it noticed precision discrepancy.

Before the change:

    clock_getres(CLOCK_MONOTONIC, &res) reported 250Hz precision.

After the change:

    clock_getres(CLOCK_MONOTONIC, &res) reports ITC (400Mhz) precision.

The patch is based on matoro's fix. I added a bit of explanation why we
need to special-case arch-specific clock_getres().

[akpm@linux-foundation.org: coding-style cleanups]
Link: https://lkml.kernel.org/r/20220820181813.2275195-1-slyich@gmail.com
Signed-off-by: Sergei Trofimovich <slyich@gmail.com>
Cc: matoro <matoro_mailinglist_kernel@matoro.tk>
Cc: Émeric Maschino <emeric.maschino@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
arch/ia64/kernel/sys_ia64.c
arch/ia64/kernel/syscalls/syscall.tbl

index e14db25146c222cdc8aef027f691fef6a73ac9cf..215bf3f8cb2043fefb26db7f0bc91c69b0440bb1 100644 (file)
@@ -166,3 +166,29 @@ ia64_mremap (unsigned long addr, unsigned long old_len, unsigned long new_len, u
                force_successful_syscall_return();
        return addr;
 }
+
+asmlinkage long
+ia64_clock_getres(const clockid_t which_clock, struct __kernel_timespec __user *tp)
+{
+       /*
+        * ia64's clock_gettime() syscall is implemented as a vdso call
+        * fsys_clock_gettime(). Currently it handles only
+        * CLOCK_REALTIME and CLOCK_MONOTONIC. Both are based on
+        * 'ar.itc' counter which gets incremented at a constant
+        * frequency. It's usually 400MHz, ~2.5x times slower than CPU
+        * clock frequency. Which is almost a 1ns hrtimer, but not quite.
+        *
+        * Let's special-case these timers to report correct precision
+        * based on ITC frequency and not HZ frequency for supported
+        * clocks.
+        */
+       switch (which_clock) {
+       case CLOCK_REALTIME:
+       case CLOCK_MONOTONIC:
+               s64 tick_ns = DIV_ROUND_UP(NSEC_PER_SEC, local_cpu_data->itc_freq);
+               struct timespec64 rtn_tp = ns_to_timespec64(tick_ns);
+               return put_timespec64(&rtn_tp, tp);
+       }
+
+       return sys_clock_getres(which_clock, tp);
+}
index 78b1d03e86e1dc1e7b71dbaaf2649d5e7e6c48cd..72c929d9902b99601ed98f25f89ed88cc2d9935e 100644 (file)
 228    common  timer_delete                    sys_timer_delete
 229    common  clock_settime                   sys_clock_settime
 230    common  clock_gettime                   sys_clock_gettime
-231    common  clock_getres                    sys_clock_getres
+231    common  clock_getres                    ia64_clock_getres
 232    common  clock_nanosleep                 sys_clock_nanosleep
 233    common  fstatfs64                       sys_fstatfs64
 234    common  statfs64                        sys_statfs64