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. It adds a bit of explanation why we
need to special-case arch-specific clock_getres().
CC: [email protected]
CC: Andrew Morton <[email protected]>
Signed-off-by: Sergei Trofimovich <[email protected]>
---
arch/ia64/kernel/sys_ia64.c | 26 ++++++++++++++++++++++++++
arch/ia64/kernel/syscalls/syscall.tbl | 2 +-
2 files changed, 27 insertions(+), 1 deletion(-)
diff --git a/arch/ia64/kernel/sys_ia64.c b/arch/ia64/kernel/sys_ia64.c
index e14db25146c2..d5d47eb4608e 100644
--- a/arch/ia64/kernel/sys_ia64.c
+++ b/arch/ia64/kernel/sys_ia64.c
@@ -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);
+}
diff --git a/arch/ia64/kernel/syscalls/syscall.tbl b/arch/ia64/kernel/syscalls/syscall.tbl
index 78b1d03e86e1..72c929d9902b 100644
--- a/arch/ia64/kernel/syscalls/syscall.tbl
+++ b/arch/ia64/kernel/syscalls/syscall.tbl
@@ -240,7 +240,7 @@
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
--
2.37.1
On Mon, 15 Aug 2022 06:49:44 +0100 Sergei Trofimovich <[email protected]> wrote:
> 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. It adds a bit of explanation why we
> need to special-case arch-specific clock_getres().
>
It would be best (and nice) to include the original developer's
Signed-off-by: and to Cc ?meric Maschino if possible?
Could you please take care of these paperwork issues?
Hi Andrew, I came up with the prototype for this patch, but it was based
entirely on Sergei's investigation which was documented in
https://bugs.gentoo.org/596382. I asked him to send it upstream because
I'm unable to attach my realname to it due to my job. I can place a
signed-off-by with my handle but I understand that's normally against
kernel policy which is why I didn't. Either way the bulk of the work
belongs to Sergei, I just scribbled it down, and he cleaned it up for
this submission.
-------- Original Message --------
Subject: Re: [PATCH] ia64: fix clock_getre(CLOCK_MONOTONIC) to report
ITC frequency
Date: 2022-08-17 15:21
From: Andrew Morton <[email protected]>
To: Sergei Trofimovich <[email protected]>
On Mon, 15 Aug 2022 06:49:44 +0100 Sergei Trofimovich <[email protected]>
wrote:
> 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. It adds a bit of explanation why we
> need to special-case arch-specific clock_getres().
>
It would be best (and nice) to include the original developer's
Signed-off-by: and to Cc Émeric Maschino if possible?
Could you please take care of these paperwork issues?
On Wed, 17 Aug 2022 12:21:03 -0700
Andrew Morton <[email protected]> wrote:
> On Mon, 15 Aug 2022 06:49:44 +0100 Sergei Trofimovich <[email protected]> wrote:
>
> > 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. It adds a bit of explanation why we
> > need to special-case arch-specific clock_getres().
> >
>
> It would be best (and nice) to include the original developer's
> Signed-off-by: and to Cc Émeric Maschino if possible?
>
> Could you please take care of these paperwork issues?
Sounds good!
I'll use matoro's nickname as is for S-o-B the way matoro is comfortable
with to share it. Will add Émeric in v2 and send out in a few minutes.
--
Sergei
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().
CC: Émeric Maschino <[email protected]>
CC: [email protected]
CC: Andrew Morton <[email protected]>
Signed-off-by: matoro <[email protected]>
Signed-off-by: Sergei Trofimovich <[email protected]>
---
Change since v1:
- No code or in-code comment change
- CCed Émeric
- Added matoro's S-O-B the way matoro comfortable with
- Fixed Subject typo s/clock_getre/clock_getres/
arch/ia64/kernel/sys_ia64.c | 26 ++++++++++++++++++++++++++
arch/ia64/kernel/syscalls/syscall.tbl | 2 +-
2 files changed, 27 insertions(+), 1 deletion(-)
diff --git a/arch/ia64/kernel/sys_ia64.c b/arch/ia64/kernel/sys_ia64.c
index e14db25146c2..d5d47eb4608e 100644
--- a/arch/ia64/kernel/sys_ia64.c
+++ b/arch/ia64/kernel/sys_ia64.c
@@ -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);
+}
diff --git a/arch/ia64/kernel/syscalls/syscall.tbl b/arch/ia64/kernel/syscalls/syscall.tbl
index 78b1d03e86e1..72c929d9902b 100644
--- a/arch/ia64/kernel/syscalls/syscall.tbl
+++ b/arch/ia64/kernel/syscalls/syscall.tbl
@@ -240,7 +240,7 @@
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
--
2.37.1
On Wed, 17 Aug 2022 17:57:42 -0400 matoro <[email protected]> wrote:
> Hi Andrew, I came up with the prototype for this patch, but it was based
> entirely on Sergei's investigation which was documented in
> https://bugs.gentoo.org/596382. I asked him to send it upstream because
> I'm unable to attach my realname to it due to my job.
In that case it's unclear that your Signed-off-by: is appropriate?
: Developer's Certificate of Origin 1.1
: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
:
: By making a contribution to this project, I certify that:
:
: (a) The contribution was created in whole or in part by me and I
: have the right to submit it under the open source license
: indicated in the file; or
:
: (b) The contribution is based upon previous work that, to the best
: of my knowledge, is covered under an appropriate open source
: license and I have the right under that license to submit that
: work with modifications, whether created in whole or in part
: by me, under the same open source license (unless I am
: permitted to submit under a different license), as indicated
: in the file; or
:
: (c) The contribution was provided directly to me by some other
: person who certified (a), (b) or (c) and I have not modified
: it.
:
: (d) I understand and agree that this project and the contribution
: are public and that a record of the contribution (including all
: personal information I submit with it, including my sign-off) is
: maintained indefinitely and may be redistributed consistent with
: this project or the open source license(s) involved.
> I can place a
> signed-off-by with my handle but I understand that's normally against
> kernel policy which is why I didn't. Either way the bulk of the work
> belongs to Sergei, I just scribbled it down, and he cleaned it up for
> this submission.
I think I'll switch it to a Cc: :)
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().
CC: Émeric Maschino <[email protected]>
CC: [email protected]
CC: Andrew Morton <[email protected]>
CC: matoro <[email protected]>
Signed-off-by: Sergei Trofimovich <[email protected]>
---
Change since v2:
- Moved matoro to CC
Change since v1:
- No code or in-code comment change
- CCed Émeric
- Added matoro's S-O-B the way matoro comfortable with
- Fixed Subject typo s/clock_getre/clock_getres/
arch/ia64/kernel/sys_ia64.c | 26 ++++++++++++++++++++++++++
arch/ia64/kernel/syscalls/syscall.tbl | 2 +-
2 files changed, 27 insertions(+), 1 deletion(-)
diff --git a/arch/ia64/kernel/sys_ia64.c b/arch/ia64/kernel/sys_ia64.c
index e14db25146c2..d5d47eb4608e 100644
--- a/arch/ia64/kernel/sys_ia64.c
+++ b/arch/ia64/kernel/sys_ia64.c
@@ -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);
+}
diff --git a/arch/ia64/kernel/syscalls/syscall.tbl b/arch/ia64/kernel/syscalls/syscall.tbl
index 78b1d03e86e1..72c929d9902b 100644
--- a/arch/ia64/kernel/syscalls/syscall.tbl
+++ b/arch/ia64/kernel/syscalls/syscall.tbl
@@ -240,7 +240,7 @@
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
--
2.37.2