Clock

NAME

clock - Kernel object used to track the progress of time.

SYNOPSIS

A clock is a one dimensional affine transformation of the clock monotonic reference timeline, which may be atomically adjusted by a clock maintainer, and observed by clients.

DESCRIPTION

Properties

The properties of a clock are established when the clock is created and cannot be changed afterwards. Currently, three clock properties are defined.

ZX_CLOCK_OPT_MONOTONIC

When set, the clock is guaranteed to have monotonic behavior. This is to say that any sequence of observations of the clock is guaranteed to produce a sequence of times that are always greater than or equal to the previous observations. A monotonic clock can never go backwards, although it can jump forwards. Formally:

Given a clock C, Let C(x) be the function that maps from the reference timeline C's timeline. C(x) is a piecewise linear function made up of all the affine transformation segments over all time as determined by C's maintainer. C is monotonic if and only if:

for all R1, R2 : R2 >= R1

C(R2) >= C(R1)

ZX_CLOCK_OPT_CONTINUOUS

When set, the clock is guaranteed to have continuous behavior. This is to say that any update to the clock transformation is guaranteed to be first order continuous with the previous transformation segment. Formally:

Let Ci(x) be the ith affine transformation segment of C(x). Let Ri be the first point in time on the reference timeline for which Ci(x) is defined. A clock C is continuous if and only if: for all i

Ci(Ri + 1) = Ci + 1(Ri + 1)

Backstop Time

The backstop time of a clock represents the minimum value that a clock may ever be set to. Since clocks can only tick forwards, and never backwards, it is impossible for an observer of a clock to ever receive a value that is less than the backstop time configured by a clock's creator.

A backstop time may be provided via the zx_create_args_v1_t structure at creation time. Otherwise, it will default to 0.

During clock update operations, any attempt to set the clock's value to something less than the backstop time will fail with ZX_ERR_INVALID_ARGS. A clock that has not been initially set will always report the backstop time configured for the clock. Backtop times may never be less than the default value of zero.

Implied properties

  • The reference clock for all clock objects in the system is clock monotonic.
  • The nominal units of all clock objects are specified to be nanoseconds. This property is not configurable.
  • The units of frequency adjustment for all clock objects are specified to be parts per million, or PPM.
  • The maximum permissible range of frequency adjustment of a clock object is specified to be [-1000, +1000] PPM. This property is not configurable.

Additional creation options

ZX_CLOCK_OPT_AUTO_START

When you use this option during clock creation, the clock begins in a started state instead of the default non-started state. See Starting a clock for details.

Reading the clock

Given a clock handle, users may query the current time given by that clock using the zx_clock_read() syscall. Clock reads ZX_RIGHT_READ permissions. Clock reads are guaranteed to be coherent for all observers. This is to say that, if two observers query the clock at exactly the same reference time R, that they will always see the same value C(R).

Reference timelines, zx_ticks_get(), and zx_clock_get_monotonic()

As noted earlier, zx_clock_get_monotonic() is the reference timeline for all user-created zircon clocks. This means that if a user knows a clock instance's current transformation, then given a value on the clock instance's timeline, the corresponding point on the clock monotonic timeline may be computed (and vice-versa). It also means that in the absence of a rate adjustment made to the kernel clock, clock monotonic and the kernel clock will tick at exactly the same rate.

In addition to the clock monotonic timeline, the zircon kernel also exposes the "ticks" timeline via zx_ticks_get() and zx_ticks_per_second(). Internally, ticks is actually the reference timeline for clock monotonic and is read directly from an architecture appropriate timer unit accessible to the kernel. Clock monotonic is actually a linear transformation of the ticks timeline normalized to nanosecond units. Both timelines start ticking from zero as the kernel starts.

Because clock monotonic is a static transformation based on ticks, and all kernel clocks are transformations based on clock monotonic, ticks may also serve as a reference clock for kernel clocks in addition to clock monotonic.

Fetching the clock's details

In addition to simply reading the current value of the clock, advanced users who possess ZX_RIGHT_READ permissions may also read the clock and get extended details in the process using zx_clock_get_details(). Upon a successful call, the details structure returned to callers will include:

  • The current clock monotonic to clock transformation.
  • The current ticks to clock transformation.
  • The current symmetric error bound estimate (if any) for the clock.
  • The last time the clock was updated as defined by the clock monotonic reference timeline.
  • An observation of the system tick counter, which was taken during the observation of the clock.
  • All of the static properties of the clock defined at creation time.
  • A generation nonce whose value changes every time the clock's underlying transformation is updated.

Advanced users may use these details to not only compute a recent now value for the clock (by transforming the reported ticks-now observation using the ticks-to-clock transformation, both reported by the get details operation), but to also:

  • Know whether the clock transformation has been changed since the last zx_clock_get_details() operation (using the generation nonce). Note that a clock's generation nonce is not guaranteed to start at any given value, nor to change in any specific way (such as incrementing by a fixed amount) with each update. Instead, with every update the generation nonce will be changed to a value which is guaranteed to be different from the value it had immediately before the update took place.
  • Compose the clock transformation with other clocks' transformations to reason about the relationship between two clocks.
  • Know the clock maintainer's best estimate of error bound.
  • Reason about the range of possible future values of the clock relative to the reference clock based on the last correction time, the current transformation, and the maximum permissible correction factor for the clock (see the maximum permissive range of frequency adjustment described in the |Implied properties| section above.

Starting a clock and clock signals

Immediately after creation, a clock has not yet been started. All attempts to read the clock will return the clock's configured backstop time, which defaults to 0 if unspecified during creation.

A clock begins running after the very first update operation performed by a clock's maintainer, which must include a set-value operation. The clock will begin running at that point with a rate equal to the reference clock plus the deviation from nominal specified by the maintainer.

Clocks also have a ZX_CLOCK_STARTED signal, which may be used by users to know when the clock has actually been started. Initially, this signal is not set, but it becomes set after the first successful update operation. Once started, a clock will never stop and the ZX_CLOCK_STARTED signal will always be asserted.

Initially, the clock is a clone of clock monotonic, which makes the transformation between the clock monotonic timeline and synthetic timeline the identity function. This clock may still be maintained after creation, subject to the limitations imposed by rights, the ZX_CLOCK_OPT_MONOTONIC and ZX_CLOCK_OPT_CONTINUOUS properties, and the configured backstop time.

If a clock is created with the ZX_CLOCK_OPT_AUTO_START options, it cannot be configured with a backstop time that is greater than the current clock monotonic time. If this was allowed, this would result in a state where a clock's current time is set to a time before its backstop time.

Maintaining a clock

Users who possess ZX_RIGHT_WRITE permissions for a clock object may act as a maintainer of the clock using the zx_clock_update() syscall. Three parameters of the clock may be adjusted during each call to zx_clock_update(), but not all three need to be adjusted each time. These values are:

  • The clock's absolute value.
  • The frequency adjustment of the clock (deviation from nominal expressed in ppm)
  • The absolute error bound estimate of the clock (expressed in nanoseconds)

Changes to a clocks transformation occur during the syscall itself. The specific reference time of the adjustment may not be specified by the user.

Any change to the absolute value of a clock with the ZX_CLOCK_OPT_MONOTONIC property set on it which would result in non-monotonic behavior will fail with a return code of ZX_ERR_INVALID_ARGS.

The first update operation is what starts a clock ticking and must include a set-value operation.

Aside from the very first set-value operation, all attempts to set the absolute value of a clock with the ZX_CLOCK_OPT_CONTINUOUS property set on it will fail with a return code of ZX_ERR_INVALID_ARGS

Notes on the clock error bound estimate

The zx_clock_get_details() syscall provides the user with a number of fine grained details about a clock, including the "error bound estimate". This value, expressed in nanoseconds, represents the clock maintainer's best current estimate of how wrong the clock currently might be relative to the reference(s) being used by the maintainer. For example, if a user fetched a time X with an error bound estimate of E, then the clock maintainer is attempting to say that it believes that the clock's actual value is somewhere in the range [ X-E, X+E ].

The level of confidence in this estimate is not specified by the kernel APIs. It is possible that some clock maintainers are using a strict bound, while others are using a bound that is not provable but provides "high confidence", while others still might have little to no confidence in their estimates.

In the case that a user needs to understand the objective quality of the error estimates they are accessing (for example, to enforce certificate validity dates, or DRM license expiration), they should understand which component in the system is maintaining their clock, and what guarantees that maintainer provides when it comes to the confidence levels of its published error bound estimates.

Mappable clocks

In addition to reading a clock or getting its details using the standard syscalls (zx_clock_read() and zx_clock_get_details()), Zircon kernel clocks offer another approach which can decrease observation overhead by (almost always) skipping the need to enter the Zircon kernel in order to observe the clock. The main idea here is that clocks are, at their core, nothing more than an affine transformation applied to the current value of the selected reference clock. If the transformation state can be accessed by user-mode via shared memory, and the underlying reference clock does not require entering the Zircon kernel to read, then the synthetic clock can be also observed without ever needing to enter the kernel. Enter "mappable" clocks.

Creation

A mappable kernel clock can be created by calling zx_clock_create() and passing ZX_CLOCK_OPT_MAPPABLE at the time of creation. The newly created clock handle will contain the ZX_RIGHT_MAP permission as well as all of the other default permissions. Additionally, the options field reported via a get-details operation will contain the ZX_CLOCK_OPT_MAPPABLE flag, reflecting the fact that this is a mappable clock.

Mapping

In order to observe the clock's state using shared memory, two things must happen first.

1) The mapped size of the clock state must be determined. 2) The clock's shared state must be mapped into the user's address space.

To fetch the clock's mapped size, users must call zx_object_get_info() with the ZX_INFO_CLOCK_MAPPED_SIZE topic, passing a uint64_t sized buffer to receive the mapped size of the clock.

Once the size is known, users may call zx_vmar_map_clock() in order to actually map the clock's state into their address space. This call operates very similarly to zx_vmap_map_vmo(), only with some extra built in restrictions. Notably:

1) Only the ZX_VM_PERM_READ permission is allowed, all of the other ZX_VM_PERM flags are are explicitly disallowed, regardless of the handle rights the user possesses. Mapped clock data must always be read only. 2) The len parameter passed to the syscall must match the value retrieved from the get-info call. 3) There is no ability to specify a vmo_offset when mapping the clock's state. The entire clock state region must always be mapped.

Mapping a clock requires both the ZX_RIGHT_READ and ZX_RIGHT_MAP permissions on the clock handle.

A successful map operation will return the virtual address of where the clock object's state has been mapped in the specified VMAR.

Unmapping

The process of unmapping a clock is identical to the process of unmapping a VMO or an IO Buffer region. Users simply call zx_vmar_unmap() on the VMAR they used to map the clock, passing the length received from the original get-info call, and virtual address received from the zx_clock_map() call.

Users may close their clock's handle after mapping it and still read the clock's state. Just like closing a VMO handle will not destroy any mappings of that VMO, closing the clock handle will also not destroy any of its mappings. See zx_vmar_unmap() and zx_vmar_destroy().

Observing a mapped clock

To observe a mapped clock, users call either zx_clock_read_mapped() or zx_clock_get_details_mapped(), passing the virtual address for the mapped clock state as the first parameter. Everything else remains the same, please refer to the documentation for zx_clock_read() and zx_clock_get_details() for more information.

Any of the following actions will result in undefined behavior:

  • Attempting to call a mapped clock observation syscall using any virtual address aside from the one returned by the original call to zx_clock_map().
  • Attempting to call a mapped clock observation syscall using clock state which has been either partially, or completely, unmapped.

Additionally, users must not make any attempt to use the mapped clock state to infer details about the clock itself. The format of the clock's state is unspecified and subject to change at any time. The only legitimate way to use a clock's mapped state is in conjunction with calls to zx_clock_read_mapped() orzx_clock_get_details_mapped()`.

When properly accessed using the provided syscalls, clock observations will be "multi-observer" consistent, meaning the properties such as monotonicity and continuity will be consistent across a set of observations made by multiple observers if those observations have an established order. See here for details.

Clock mapping representation in zx_info_maps_t and zx_info_vmos_t structures

Zircon currently provides two diagnostic zx_object_get_info() topics for fetching information about about currently active mappings, ZX_INFO_VMAR_MAPS and ZX_INFO_PROCESS_MAPS, both of which will return an array of zx_info_maps_t strcutres describing the currently active mappings in the specified VMAR, or process.

Clock mappings will be enumerated in the same way that VMO mappings are, with the following differences.

1) Since clocks are not currently nameable objects, the string "kernel-clock" will be returned as the name of the mapping, instead of the current name of a VMO as would typically be done. 2) The KOID field of the mapping will be populated using the clock object's KOID.

Additionally, programs can get info on the current VMOs associated with a process using the ZX_INFO_PROCESS_VMOS topic to fetch an array of zx_info_vmos_t structures. When a process has a clock mapped into its address space, the clock will show up in this enumaration as well. Just like in the zx_info_maps_t case, the KOID reported will be the KOID of the clock object, and the name reported will be "kernel-clock". Additionally, the flags field of the entry will have the ZX_INFO_VMO_VIA_MAPPING bit set to indicate that the object is being enumerated on the list because of the active mapping.

Example

Here is a simplified example showing how you might map and read a clock.

class MappedClockReader {
 public:
  MappedClockReader(const zx::clock& clock) {
    zx_status_t status = clock.get_info(ZX_INFO_CLOCK_MAPPED_SIZE, &mapped_clock_size_,
                                        sizeof(mapped_clock_size_), nullptr, nullptr));
    ZX_ASSERT(status == ZX_OK);

    status = zx::vmar::root_self()->map_clock(ZX_VM_PERM_READ | ZX_VM_MAP_RANGE, 0, clock,
                                              mapped_clock_size_, &clock_addr_);
    ZX_ASSERT(status == ZX_OK);
  }

  ~MappedClockReader() {
    if (clock_addr_ != 0) {
      zx::vmar::root_self()->unmap(clock_addr_, mapped_clock_size_);
    }
  }

  zx_time_t Read() {
    zx_time_t result{0};
    const zx_status_t status = zx::clock::read_mapped(clock_addr_, &result);
    ZX_ASSERT(status == ZX_OK);
    return result;
  }

 private:
  zx_vaddr_t clock_addr_{0};
  uint64_t mapped_clock_size_{0};
}

REFERENCES