The comment above bintime2timespec() says:
When converting between timestamps on parallel timescales of differing
resolutions it is historical and scientific practice to round down.
However, the delta_nsec value is a time difference and not a timestamp. Also
the rounding errors accumulate in the frequency accumulator, see hardpps().
So, rounding to the closest integer is probably slightly better.
Reviewed by: imp
Pull Request: https://github.com/freebsd/freebsd-src/pull/604
Let A be the current calculation of the frequency accumulator (pps_fcount)
update in pps_event()
scale = (uint64_t)1 << 63;
scale /= captc->tc_frequency;
scale *= 2;
bt.sec = 0;
bt.frac = 0;
bintime_addx(&bt, scale * tcount);
bintime2timespec(&bt, &ts);
hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
and hardpps(..., delta_nsec):
u_nsec = delta_nsec;
if (u_nsec > (NANOSECOND >> 1))
u_nsec -= NANOSECOND;
else if (u_nsec < -(NANOSECOND >> 1))
u_nsec += NANOSECOND;
pps_fcount += u_nsec;
This change introduces a new calculation which is slightly simpler and more
straight forward. Name it B.
Consider the following sample values with a tcount of 2000000100 and a
tc_frequency of 2000000000 (2GHz).
For A, the scale is 9223372036. Then scale * tcount is 18446744994337203600
which is larger than UINT64_MAX (= 18446744073709551615). The result is
920627651984 == 18446744994337203600 % UINT64_MAX. Since all operands are
unsigned the result is well defined through modulo arithmetic. The result of
bintime2timespec(&bt, &ts) is 49. This is equal to the correct result
1000000049 % NANOSECOND.
In hardpps(), both conditional statements are not executed and pps_fcount is
incremented by 49.
For the new calculation B, we have 1000000000 * tcount is 2000000100000000000
which is less than UINT64_MAX. This yields after the division with tc_frequency
the correct result of 1000000050 for delta_nsec.
In hardpps(), the first conditional statement is executed and pps_fcount is
incremented by 50.
This shows that both methods yield roughly the same results. However, method B
is easier to understand and requires fewer conditional statements.
Reviewed by: imp
Pull Request: https://github.com/freebsd/freebsd-src/pull/604
Use local variables for the captured timehand and timecounter in pps_event().
This fixes a potential issue in the nsec preparation for hardpps(). Here the
timecounter was accessed through the captured timehand after the generation was
checked.
Make a snapshot of the relevent timehand values early in pps_event(). Check
the timehand generation only once during the capture and event processing. Use
atomic_thread_fence_acq() similar to the other readers.
Reviewed by: imp
Pull Request: https://github.com/freebsd/freebsd-src/pull/604
The third argument to this function indicates whether the supplied
ticker is fixed or variable, i.e. requiring calibration. Give this
argument a type and name that better conveys this purpose.
Reviewed by: kib, markj
MFC after: 1 week
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D35459
cpu ticks) has some imprecision and, worse, huge timestep (about
20 minutes on 4GHz CPU) near 53.4 days of elapsed time.
kern_time.c/cputick2timespec() (it is used for clock_gettime() for
querying process or thread consumed cpu time) Uses cputick2usec()
and then needlessly converting usec to nsec, obviously losing
precision even with fixed cputick2usec().
kern_time.c/kern_clock_getres() uses some weird (anyway wrong)
formula for getting cputick resolution.
PR: 262215
Reviewed by: gnn
Differential Revision: https://reviews.freebsd.org/D34558
There are two places where we convert from a timecounter delta to
a bintime delta: tc_windup and bintime_off.
Both functions use the same calculations when the timecounter delta is
small. But for a large delta (greater than approximately an equivalent
of 1 second) the calculations were different. Both functions use
approximate calculations based on th_scale that avoid division. Both
produce values slightly greater than a true value, calculated with
division by tc_frequency, would be. tc_windup is slightly more
accurate, so its result is closer to the true value and, thus, smaller
than bintime_off result.
As a consequence there can be a jump back in time when time hands are
switched after a long period of time (a large delta). Just before the
switch the time would be calculated with a large delta from
th_offset_count in bintime_off. tc_windup does the switch using its own
calculations of a new th_offset using the large delta. As explained
earlier, the new th_offset may end up being less than the previously
produced binuptime. So, for a period of time new binuptime values may
be "back in time" comparing to values just before the switch.
Such a jump must never happen. All the code assumes that the uptime is
monotonically nondecreasing and some code works incorrectly when that
assumption is broken. For example, we have observed sleepq_timeout()
ignoring a timeout when the sbinuptime value obtained by the callout
code was greater than the expiration value, but the sbinuptime obtained
in sleepq_timeout() was less than it. In that case the target thread
would never get woken up.
The unified calculations should ensure the monotonic property of the
uptime.
The problem is quite rare as normally tc_windup should be called HZ
times per second (typically 1000 or 100). But it may happen in VMs on
very busy hypervisors where a VM's virtual CPU may not get an execution
time slot for a second or more.
Reviewed by: kib
MFC after: 2 weeks
Sponsored by: Panzura LLC
Hyper-V wants to register its MSR-based timecounter during
SI_SUB_HYPERVISOR, before SI_SUB_LOCK, since an emulated 8254 may not be
available for DELAY(). So we cannot use MTX_SYSINIT to initialize the
timecounter lock.
PR: 259878
Reviewed by: kib
MFC after: 3 days
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D33014
This change is a slight performance optimization for systems with a slow
64-bit division.
The th->th_scale and th->th_large_delta values only depend on the
timecounter frequency and the th->th_adjustment. The timecounter
frequency of a timehand only changes when a new timecounter is activated
for the timehand. The th->th_adjustment is only changed by the NTP
second update. The NTP second update is not done for every call of
tc_windup().
Move the code block to recalculate the scaling factor and
the large delta of a timehand to the new helper function
recalculate_scaling_factor_and_large_delta().
Call recalculate_scaling_factor_and_large_delta() when a new timecounter
is activated and a NTP second update occurred.
MFC after: 1 week
Timecounter registration is dynamic, i.e., there is no requirement that
timecounters must be registered during single-threaded boot. Loadable
drivers may in principle register timecounters (which can be switched to
automatically). Timecounters cannot be unregistered, though this could
be implemented.
Registered timecounters belong to a global linked list. Add a mutex to
synchronize insertions and the traversals done by (mpsafe) sysctl
handlers. No functional change intended.
Reviewed by: imp, kib
MFC after: 2 weeks
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D32511
These ones were unambiguous cases where the Foundation was the only
listed copyright holder (in the associated license block).
Sponsored by: The FreeBSD Foundation
In FreeBSD, the current time is computed from uptime + boottime. Uptime
is a continuous, smooth function that's monotonically increasing. To
effect changes to the current time, boottime is adjusted. boottime is
mutable and shouldn't be cached against future need. Document the
current implementation, with the caveat that we may stop stepping
boottime on resume in the future and will step uptime instead (noted in
the commit message, but not in the code).
Sponsored by: Netflix
Reviewed by: phk, rpokala
Differential Revision: https://reviews.freebsd.org/D30116
- Implement a dtrace_getnanouptime(), matching the existing
dtrace_getnanotime(), to avoid DTrace calling out to a potentially
instrumentable function.
(These should probably both be under KDTRACE_HOOKS. Also, it's not clear
to me that they are correct implementations for the DTrace thread time
functions they are used in .. fixes for another commit.)
- Don't allow FBT to instrument functions involved in EL1 exception handling
that are involved in FBT trap processing: handle_el1h_sync() and
do_el1h_sync().
- Don't allow FBT to instrument DDB and KDB functions, as that makes it
rather harder to debug FBT problems.
Prior to these changes, use of FBT on FreeBSD/arm64 rapidly led to kernel
panics due to recursion in DTrace.
Reliable FBT on FreeBSD/arm64 is reliant on another change from @andrew to
have the aarch64 instrumentor more carefully check that instructions it
replaces are against the stack pointer, which can otherwise lead to memory
corruption. That change remains under review.
MFC after: 2 weeks
Reviewed by: andrew, kp, markj (earlier version), jrtc27 (earlier version)
Differential revision: https://reviews.freebsd.org/D27766
It seems that second call does not add any useful state change for all
implemented timecounters.
Discussed with: bde
Sponsored by: The FreeBSD Foundation
MFC after: 3 weeks
r357614 added CTLFLAG_NEEDGIANT to make it easier to find nodes that are
still not MPSAFE (or already are but aren’t properly marked).
Use it in preparation for a general review of all nodes.
This is non-functional change that adds annotations to SYSCTL_NODE and
SYSCTL_PROC nodes using one of the soon-to-be-required flags.
Mark all obvious cases as MPSAFE. All entries that haven't been marked
as MPSAFE before are by default marked as NEEDGIANT
Approved by: kib (mentor, blanket)
Commented by: kib, gallatin, melifaro
Differential Revision: https://reviews.freebsd.org/D23718
bintime()/binuptime().
The algorithm to read the consistent snapshot of current timehand is
repeated in each accessor, including the details proper rollup
detection and synchronization with the writer. In fact there are only
two different kind of readers: one for bintime()/binuptime() which has
to do the in-place calculation, and another kind which fetches some
member from struct timehand.
Extract the logic into type-checked macros, GETTHBINTIME() for bintime
calculation, and GETTHMEMBER() for safe read of a structure' member.
This way, the synchronization is only written in bintime_off() and
getthmember().
In bintime_off(), use overflow-safe calculation of th_scale *
delta(timecounter). In tc_windup, pre-calculate the min delta value
which overflows and require slow algorithm, into the new timehands
th_large_delta member.
This part with overflow fix was written by Bruce Evans.
Reported by: Mark Millard <marklmi@yahoo.com> (the overflow issue)
Tested by: pho
Discussed with: emaste
Sponsored by: The FreeBSD Foundation (kib)
MFC after: 3 weeks
The timespecadd(3) family of macros were imported from NetBSD back in
r35029. However, they were initially guarded by #ifdef _KERNEL. In the
meantime, we have grown at least 28 syscalls that use timespecs in some
way, leading many programs both inside and outside of the base system to
redefine those macros. It's better just to make the definitions public.
Our kernel currently defines two-argument versions of timespecadd and
timespecsub. NetBSD, OpenBSD, and FreeDesktop.org's libbsd, however, define
three-argument versions. Solaris also defines a three-argument version, but
only in its kernel. This revision changes our definition to match the
common three-argument version.
Bump _FreeBSD_version due to the breaking KPI change.
Discussed with: cem, jilles, ian, bde
Differential Revision: https://reviews.freebsd.org/D14725
On arm64 (and possible other architectures) we are unable to use static
DPCPU data in kernel modules. This is because the compiler will generate
PC-relative accesses, however the runtime-linker expects to be able to
relocate these.
In preparation to fix this create two macros depending on if the data is
global or static.
Reviewed by: bz, emaste, markj
Sponsored by: ABT Systems Ltd
Differential Revision: https://reviews.freebsd.org/D16140
opt_compat.h is mentioned in nearly 180 files. In-progress network
driver compabibility improvements may add over 100 more so this is
closer to "just about everywhere" than "only some files" per the
guidance in sys/conf/options.
Keep COMPAT_LINUX32 in opt_compat.h as it is confined to a subset of
sys/compat/linux/*.c. A fake _COMPAT_LINUX option ensure opt_compat.h
is created on all architectures.
Move COMPAT_LINUXKPI to opt_dontuse.h as it is only used to control the
set of compiled files.
Reviewed by: kib, cem, jhb, jtl
Sponsored by: DARPA, AFRL
Differential Revision: https://reviews.freebsd.org/D14941
In this case volatile qualifiers enusre that a compiler does not
optimize the accesses out.
Reviewed by: alc, jhb
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
Differential revision: https://reviews.freebsd.org/D13534
all cache the last system time (uptime + boottime). Only the format
differs. Do not re-calculate the bintime and simply use the value
used to calculate the microtime and nanotime.
Group all the updates under the relevant comment. Remove obsoleted
XXX part.
Submitted by: Sebastian Huber <sebastian.huber@embedded-brains.de>
MFC after: 1 week
I moved this branch from github to a private server, and pulled from the
wrong one when committing r315280, so I failed to include two recent commits.
Thankfully, they were only cosmetic and were included in the review.
Specifically:
Add documentation, polish comments, and improve style(9).
Tested by: pho (r315280)
MFC after: 2 weeks
Sponsored by: Dell EMC
Differential Revision: https://reviews.freebsd.org/D9791
POSIX 2008 says this about clock_settime(2):
If the value of the CLOCK_REALTIME clock is set via clock_settime(),
the new value of the clock shall be used to determine the time
of expiration for absolute time services based upon the
CLOCK_REALTIME clock. This applies to the time at which armed
absolute timers expire. If the absolute time requested at the
invocation of such a time service is before the new value of
the clock, the time service shall expire immediately as if the
clock had reached the requested time normally.
Setting the value of the CLOCK_REALTIME clock via clock_settime()
shall have no effect on threads that are blocked waiting for
a relative time service based upon this clock, including the
nanosleep() function; nor on the expiration of relative timers
based upon this clock. Consequently, these time services shall
expire when the requested relative interval elapses, independently
of the new or old value of the clock.
When the real-time clock is adjusted, such as by clock_settime(3),
wake any threads sleeping until an absolute real-clock time.
Such a sleep is indicated by a non-zero td_rtcgen. The sleep functions
will set that field to zero and return zero to tell the caller
to reevaluate its sleep duration based on the new value of the clock.
At present, this affects the following functions:
pthread_cond_timedwait(3)
pthread_mutex_timedlock(3)
pthread_rwlock_timedrdlock(3)
pthread_rwlock_timedwrlock(3)
sem_timedwait(3)
sem_clockwait_np(3)
I'm working on adding clock_nanosleep(2), which will also be affected.
Reported by: Sebastian Huber <sebastian.huber@embedded-brains.de>
Reviewed by: jhb, kib
MFC after: 2 weeks
Relnotes: yes
Sponsored by: Dell EMC
Differential Revision: https://reviews.freebsd.org/D9791
Sysctls like kern.eventtimer.et.*.quality currently embed the name of
the clock device. This is problematic for the Prometheus metrics
exporter for two reasons:
- Some of those clocks have dashes in their names, which Prometheus
doesn't allow to be used in metric names.
- It doesn't allow for extracting the same property of all clocks on the
system from within a single query.
Attach these nodes to have a label, so that the Prometheus metrics
exporter gives these metric a uniform name with the name of the clock
attached as a label.
Reviewed by: cem
Differential Revision: https://reviews.freebsd.org/D8775
Right now, userspace (fast) gettimeofday(2) on x86 only works for
RDTSC. For older machines, like Core2, where RDTSC is not C2/C3
invariant, and which fall to HPET hardware, this means that the call
has both the penalty of the syscall and of the uncached hw behind the
QPI or PCIe connection to the sought bridge. Nothing can me done
against the access latency, but the syscall overhead can be removed.
System already provides mappable /dev/hpetX devices, which gives
straight access to the HPET registers page.
Add yet another algorithm to the x86 'vdso' timehands. Libc is updated
to handle both RDTSC and HPET. For HPET, the index of the hpet device
to mmap is passed from kernel to userspace, index might be changed and
libc invalidates its mapping as needed.
Remove cpu_fill_vdso_timehands() KPI, instead require that
timecounters which can be used from userspace, to provide
tc_fill_vdso_timehands{,32}() methods. Merge i386 and amd64
libc/<arch>/sys/__vdso_gettc.c into one source file in the new
libc/x86/sys location. __vdso_gettc() internal interface is changed
to move timecounter algorithm detection into the MD code.
Measurements show that RDTSC even with the syscall overhead is faster
than userspace HPET access. But still, userspace HPET is three-four
times faster than syscall HPET on several Core2 and SandyBridge
machines.
Tested by: Howard Su <howard0su@gmail.com>
Sponsored by: The FreeBSD Foundation
MFC after: 1 month
Differential revision: https://reviews.freebsd.org/D7473
setclock() and from simultaneous top-level and interrupt. For this,
tc_windup() is protected with a tc_setclock_mtx spinlock, in the try
mode when called from hardclock interrupt. If spinlock cannot be
obtained without spinning from the interrupt context, this means that
top-level executes tc_windup() on other core and our try may be
avoided.
The boottimebin and boottime variables should be adjusted from
tc_windup(). To be correct, they must be part of the timehands and
read using lockless protocol. Remove the globals and reimplement the
getboottime(9)/getboottimebin(9) KPI using the timehands read
protocol.
Tested by: pho (as part of the whole patch)
Reviewed by: jhb (same)
Discussed wit: bde
Sponsored by: The FreeBSD Foundation
MFC after: 1 month
X-Differential revision: https://reviews.freebsd.org/D7302
consumers can now be only one tc_windup() call late.
Use C99 initialization.
Tested by: pho (as part of the whole patch)
Reviewed by: jhb (same)
Discussed with: bde
Sponsored by: The FreeBSD Foundation
MFC after: 1 month
X-Differential revision: https://reviews.freebsd.org/D7302
and getboottimebin(9) KPI. Change consumers of boottime to use the
KPI. The variables were renamed to avoid shadowing issues with local
variables of the same name.
Issue is that boottime* should be adjusted from tc_windup(), which
requires them to be members of the timehands structure. As a
preparation, this commit only introduces the interface.
Some uses of boottime were found doubtful, e.g. NLM uses boottime to
identify the system boot instance. Arguably the identity should not
change on the leap second adjustment, but the commit is about the
timekeeping code and the consumers were kept bug-to-bug compatible.
Tested by: pho (as part of the bigger patch)
Reviewed by: jhb (same)
Discussed with: bde
Sponsored by: The FreeBSD Foundation
MFC after: 1 month
X-Differential revision: https://reviews.freebsd.org/D7302
are updated lockess, different CPUs write its own view of timecounter
state. The critical section is done for safety, callers of
tc_cpu_ticks() are supposed to already enter critical section, or to
own a spinlock.
The change fixes sporadical reports of too high values reported for
the (W)CPU on platforms that do not provide cpu ticker and use
tc_cpu_ticks(), in particular, arm*.
Diagnosed and reviewed by: jhb
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
with higher quality registers (presumably in a module that has just been
loaded), do not undo the user's choice by switching to the new timecounter.
Document that behavior, and also the fact that there is no way to unregister
a timecounter (and thus no way to unload a module containing one).
This makes the PPS API behave correctly, but isn't ideal -- we still end
up capturing PPS data for non-enabled edges, we just don't process the
data into an event that becomes visible outside of kern_tc. That's because
the event type isn't passed to pps_capture(), so it can't do the filtering.
Any solution for capture filtering is going to require touching every driver.
for timehands consumers, by using fences.
Ensure that the timehands->th_generation reset to zero is visible
before the data update is visible [*]. tc_setget() allowed data update
writes to become visible before generation (but not on TSO
architectures).
Remove tc_setgen(), tc_getgen() helpers, use atomics inline [**].
Noted by: alc [*]
Requested by: bde [**]
Reviewed by: alc, bde
Sponsored by: The FreeBSD Foundation
MFC after: 3 weeks
Do not include machine/atomic.h explicitely, the header is already included
by sys/systm.h.
Force inlining of tc_getgen() and tc_setgen(). The functions are used
more than once, which causes compilers with non-aggressive inlining
policies to generate calls.
Suggested by: bde
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
