1/*
2 * Copyright (C) 2011 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 *      http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17/*
18 * A service that exchanges time synchronization information between
19 * a master that defines a timeline and clients that follow the timeline.
20 */
21
22#define __STDC_LIMIT_MACROS
23#define LOG_TAG "common_time"
24#include <utils/Log.h>
25#include <stdint.h>
26
27#include <common_time/local_clock.h>
28#include <assert.h>
29
30#include "clock_recovery.h"
31#include "common_clock.h"
32#ifdef TIME_SERVICE_DEBUG
33#include "diag_thread.h"
34#endif
35
36// Define log macro so we can make LOGV into LOGE when we are exclusively
37// debugging this code.
38#ifdef TIME_SERVICE_DEBUG
39#define LOG_TS ALOGE
40#else
41#define LOG_TS ALOGV
42#endif
43
44namespace android {
45
46ClockRecoveryLoop::ClockRecoveryLoop(LocalClock* local_clock,
47                                     CommonClock* common_clock) {
48    assert(NULL != local_clock);
49    assert(NULL != common_clock);
50
51    local_clock_  = local_clock;
52    common_clock_ = common_clock;
53
54    local_clock_can_slew_ = local_clock_->initCheck() &&
55                           (local_clock_->setLocalSlew(0) == OK);
56    tgt_correction_ = 0;
57    cur_correction_ = 0;
58
59    // Precompute the max rate at which we are allowed to change the VCXO
60    // control.
61    uint64_t N = 0x10000ull * 1000ull;
62    uint64_t D = local_clock_->getLocalFreq() * kMinFullRangeSlewChange_mSec;
63    LinearTransform::reduce(&N, &D);
64    while ((N > INT32_MAX) || (D > UINT32_MAX)) {
65        N >>= 1;
66        D >>= 1;
67        LinearTransform::reduce(&N, &D);
68    }
69    time_to_cur_slew_.a_to_b_numer = static_cast<int32_t>(N);
70    time_to_cur_slew_.a_to_b_denom = static_cast<uint32_t>(D);
71
72    reset(true, true);
73
74#ifdef TIME_SERVICE_DEBUG
75    diag_thread_ = new DiagThread(common_clock_, local_clock_);
76    if (diag_thread_ != NULL) {
77        status_t res = diag_thread_->startWorkThread();
78        if (res != OK)
79            ALOGW("Failed to start A@H clock recovery diagnostic thread.");
80    } else
81        ALOGW("Failed to allocate diagnostic thread.");
82#endif
83}
84
85ClockRecoveryLoop::~ClockRecoveryLoop() {
86#ifdef TIME_SERVICE_DEBUG
87    diag_thread_->stopWorkThread();
88#endif
89}
90
91// Constants.
92const float ClockRecoveryLoop::dT = 1.0;
93const float ClockRecoveryLoop::Kc = 1.0f;
94const float ClockRecoveryLoop::Ti = 15.0f;
95const float ClockRecoveryLoop::Tf = 0.05;
96const float ClockRecoveryLoop::bias_Fc = 0.01;
97const float ClockRecoveryLoop::bias_RC = (dT / (2 * 3.14159f * bias_Fc));
98const float ClockRecoveryLoop::bias_Alpha = (dT / (bias_RC + dT));
99const int64_t ClockRecoveryLoop::panic_thresh_ = 50000;
100const int64_t ClockRecoveryLoop::control_thresh_ = 10000;
101const float ClockRecoveryLoop::COmin = -100.0f;
102const float ClockRecoveryLoop::COmax = 100.0f;
103const uint32_t ClockRecoveryLoop::kMinFullRangeSlewChange_mSec = 300;
104const int ClockRecoveryLoop::kSlewChangeStepPeriod_mSec = 10;
105
106
107void ClockRecoveryLoop::reset(bool position, bool frequency) {
108    Mutex::Autolock lock(&lock_);
109    reset_l(position, frequency);
110}
111
112uint32_t ClockRecoveryLoop::findMinRTTNdx(DisciplineDataPoint* data,
113                                          uint32_t count) {
114    uint32_t min_rtt = 0;
115    for (uint32_t i = 1; i < count; ++i)
116        if (data[min_rtt].rtt > data[i].rtt)
117            min_rtt = i;
118
119    return min_rtt;
120}
121
122bool ClockRecoveryLoop::pushDisciplineEvent(int64_t local_time,
123                                            int64_t nominal_common_time,
124                                            int64_t rtt) {
125    Mutex::Autolock lock(&lock_);
126
127    int64_t local_common_time = 0;
128    common_clock_->localToCommon(local_time, &local_common_time);
129    int64_t raw_delta = nominal_common_time - local_common_time;
130
131#ifdef TIME_SERVICE_DEBUG
132    ALOGE("local=%lld, common=%lld, delta=%lld, rtt=%lld\n",
133         local_common_time, nominal_common_time,
134         raw_delta, rtt);
135#endif
136
137    // If we have not defined a basis for common time, then we need to use these
138    // initial points to do so.  In order to avoid significant initial error
139    // from a particularly bad startup data point, we collect the first N data
140    // points and choose the best of them before moving on.
141    if (!common_clock_->isValid()) {
142        if (startup_filter_wr_ < kStartupFilterSize) {
143            DisciplineDataPoint& d =  startup_filter_data_[startup_filter_wr_];
144            d.local_time = local_time;
145            d.nominal_common_time = nominal_common_time;
146            d.rtt = rtt;
147            startup_filter_wr_++;
148        }
149
150        if (startup_filter_wr_ == kStartupFilterSize) {
151            uint32_t min_rtt = findMinRTTNdx(startup_filter_data_,
152                    kStartupFilterSize);
153
154            common_clock_->setBasis(
155                    startup_filter_data_[min_rtt].local_time,
156                    startup_filter_data_[min_rtt].nominal_common_time);
157        }
158
159        return true;
160    }
161
162    int64_t observed_common;
163    int64_t delta;
164    float delta_f, dCO;
165    int32_t tgt_correction;
166
167    if (OK != common_clock_->localToCommon(local_time, &observed_common)) {
168        // Since we just checked to make certain that this conversion was valid,
169        // and no one else in the system should be messing with it, if this
170        // conversion is suddenly invalid, it is a good reason to panic.
171        ALOGE("Failed to convert local time to common time in %s:%d",
172                __PRETTY_FUNCTION__, __LINE__);
173        return false;
174    }
175
176    // Implement a filter which should match NTP filtering behavior when a
177    // client is associated with only one peer of lower stratum.  Basically,
178    // always use the best of the N last data points, where best is defined as
179    // lowest round trip time.  NTP uses an N of 8; we use a value of 6.
180    //
181    // TODO(johngro) : experiment with other filter strategies.  The goal here
182    // is to mitigate the effects of high RTT data points which typically have
183    // large asymmetries in the TX/RX legs.  Downside of the existing NTP
184    // approach (particularly because of the PID controller we are using to
185    // produce the control signal from the filtered data) are that the rate at
186    // which discipline events are actually acted upon becomes irregular and can
187    // become drawn out (the time between actionable event can go way up).  If
188    // the system receives a strong high quality data point, the proportional
189    // component of the controller can produce a strong correction which is left
190    // in place for too long causing overshoot.  In addition, the integral
191    // component of the system currently is an approximation based on the
192    // assumption of a more or less homogeneous sampling of the error.  Its
193    // unclear what the effect of undermining this assumption would be right
194    // now.
195
196    // Two ideas which come to mind immediately would be to...
197    // 1) Keep a history of more data points (32 or so) and ignore data points
198    //    whose RTT is more than a certain number of standard deviations outside
199    //    of the norm.
200    // 2) Eliminate the PID controller portion of this system entirely.
201    //    Instead, move to a system which uses a very wide filter (128 data
202    //    points or more) with a sum-of-least-squares line fitting approach to
203    //    tracking the long term drift.  This would take the place of the I
204    //    component in the current PID controller.  Also use a much more narrow
205    //    outlier-rejector filter (as described in #1) to drive a short term
206    //    correction factor similar to the P component of the PID controller.
207    assert(filter_wr_ < kFilterSize);
208    filter_data_[filter_wr_].local_time           = local_time;
209    filter_data_[filter_wr_].observed_common_time = observed_common;
210    filter_data_[filter_wr_].nominal_common_time  = nominal_common_time;
211    filter_data_[filter_wr_].rtt                  = rtt;
212    filter_data_[filter_wr_].point_used           = false;
213    uint32_t current_point = filter_wr_;
214    filter_wr_ = (filter_wr_ + 1) % kFilterSize;
215    if (!filter_wr_)
216        filter_full_ = true;
217
218    uint32_t scan_end = filter_full_ ? kFilterSize : filter_wr_;
219    uint32_t min_rtt = findMinRTTNdx(filter_data_, scan_end);
220    // We only use packets with low RTTs for control. If the packet RTT
221    // is less than the panic threshold, we can probably eat the jitter with the
222    // control loop. Otherwise, take the packet only if it better than all
223    // of the packets we have in the history. That way we try to track
224    // something, even if it is noisy.
225    if (current_point == min_rtt || rtt < control_thresh_) {
226        delta_f = delta = nominal_common_time - observed_common;
227
228        last_error_est_valid_ = true;
229        last_error_est_usec_ = delta;
230
231        // Compute the error then clamp to the panic threshold.  If we ever
232        // exceed this amt of error, its time to panic and reset the system.
233        // Given that the error in the measurement of the error could be as
234        // high as the RTT of the data point, we don't actually panic until
235        // the implied error (delta) is greater than the absolute panic
236        // threashold plus the RTT.  IOW - we don't panic until we are
237        // absoluely sure that our best case sync is worse than the absolute
238        // panic threshold.
239        int64_t effective_panic_thresh = panic_thresh_ + rtt;
240        if ((delta > effective_panic_thresh) ||
241            (delta < -effective_panic_thresh)) {
242            // PANIC!!!
243            reset_l(false, true);
244            return false;
245        }
246
247    } else {
248        // We do not have a good packet to look at, but we also do not want to
249        // free-run the clock at some crazy slew rate. So we guess the
250        // trajectory of the clock based on the last controller output and the
251        // estimated bias of our clock against the master.
252        // The net effect of this is that CO == CObias after some extended
253        // period of no feedback.
254        delta_f = last_delta_f_ - dT*(CO - CObias);
255        delta = delta_f;
256    }
257
258    // Velocity form PI control equation.
259    dCO = Kc * (1.0f + dT/Ti) * delta_f - Kc * last_delta_f_;
260    CO += dCO * Tf; // Filter CO by applying gain <1 here.
261
262    // Save error terms for later.
263    last_delta_f_ = delta_f;
264
265    // Clamp CO to +/- 100ppm.
266    if (CO < COmin)
267        CO = COmin;
268    else if (CO > COmax)
269        CO = COmax;
270
271    // Update the controller bias.
272    CObias = bias_Alpha * CO + (1.0f - bias_Alpha) * lastCObias;
273    lastCObias = CObias;
274
275    // Convert PPM to 16-bit int range. Add some guard band (-0.01) so we
276    // don't get fp weirdness.
277    tgt_correction = CO * 327.66;
278
279    // If there was a change in the amt of correction to use, update the
280    // system.
281    setTargetCorrection_l(tgt_correction);
282
283    LOG_TS("clock_loop %lld %f %f %f %d\n", raw_delta, delta_f, CO, CObias, tgt_correction);
284
285#ifdef TIME_SERVICE_DEBUG
286    diag_thread_->pushDisciplineEvent(
287            local_time,
288            observed_common,
289            nominal_common_time,
290            tgt_correction,
291            rtt);
292#endif
293
294    return true;
295}
296
297int32_t ClockRecoveryLoop::getLastErrorEstimate() {
298    Mutex::Autolock lock(&lock_);
299
300    if (last_error_est_valid_)
301        return last_error_est_usec_;
302    else
303        return ICommonClock::kErrorEstimateUnknown;
304}
305
306void ClockRecoveryLoop::reset_l(bool position, bool frequency) {
307    assert(NULL != common_clock_);
308
309    if (position) {
310        common_clock_->resetBasis();
311        startup_filter_wr_ = 0;
312    }
313
314    if (frequency) {
315        last_error_est_valid_ = false;
316        last_error_est_usec_ = 0;
317        last_delta_f_ = 0.0;
318        CO = 0.0f;
319        lastCObias = CObias = 0.0f;
320        setTargetCorrection_l(0);
321        applySlew_l();
322    }
323
324    filter_wr_   = 0;
325    filter_full_ = false;
326}
327
328void ClockRecoveryLoop::setTargetCorrection_l(int32_t tgt) {
329    // When we make a change to the slew rate, we need to be careful to not
330    // change it too quickly as it can anger some HDMI sinks out there, notably
331    // some Sony panels from the 2010-2011 timeframe.  From experimenting with
332    // some of these sinks, it seems like swinging from one end of the range to
333    // another in less that 190mSec or so can start to cause trouble.  Adding in
334    // a hefty margin, we limit the system to a full range sweep in no less than
335    // 300mSec.
336    if (tgt_correction_ != tgt) {
337        int64_t now = local_clock_->getLocalTime();
338        status_t res;
339
340        tgt_correction_ = tgt;
341
342        // Set up the transformation to figure out what the slew should be at
343        // any given point in time in the future.
344        time_to_cur_slew_.a_zero = now;
345        time_to_cur_slew_.b_zero = cur_correction_;
346
347        // Make sure the sign of the slope is headed in the proper direction.
348        bool needs_increase = (cur_correction_ < tgt_correction_);
349        bool is_increasing  = (time_to_cur_slew_.a_to_b_numer > 0);
350        if (( needs_increase && !is_increasing) ||
351            (!needs_increase &&  is_increasing)) {
352            time_to_cur_slew_.a_to_b_numer = -time_to_cur_slew_.a_to_b_numer;
353        }
354
355        // Finally, figure out when the change will be finished and start the
356        // slew operation.
357        time_to_cur_slew_.doReverseTransform(tgt_correction_,
358                                             &slew_change_end_time_);
359
360        applySlew_l();
361    }
362}
363
364bool ClockRecoveryLoop::applySlew_l() {
365    bool ret = true;
366
367    // If cur == tgt, there is no ongoing sleq rate change and we are already
368    // finished.
369    if (cur_correction_ == tgt_correction_)
370        goto bailout;
371
372    if (local_clock_can_slew_) {
373        int64_t now = local_clock_->getLocalTime();
374        int64_t tmp;
375
376        if (now >= slew_change_end_time_) {
377            cur_correction_ = tgt_correction_;
378            next_slew_change_timeout_.setTimeout(-1);
379        } else {
380            time_to_cur_slew_.doForwardTransform(now, &tmp);
381
382            if (tmp > INT16_MAX)
383                cur_correction_ = INT16_MAX;
384            else if (tmp < INT16_MIN)
385                cur_correction_ = INT16_MIN;
386            else
387                cur_correction_ = static_cast<int16_t>(tmp);
388
389            next_slew_change_timeout_.setTimeout(kSlewChangeStepPeriod_mSec);
390            ret = false;
391        }
392
393        local_clock_->setLocalSlew(cur_correction_);
394    } else {
395        // Since we are not actually changing the rate of a HW clock, we don't
396        // need to worry to much about changing the slew rate so fast that we
397        // anger any downstream HDMI devices.
398        cur_correction_ = tgt_correction_;
399        next_slew_change_timeout_.setTimeout(-1);
400
401        // The SW clock recovery implemented by the common clock class expects
402        // values expressed in PPM. CO is in ppm.
403        common_clock_->setSlew(local_clock_->getLocalTime(), CO);
404    }
405
406bailout:
407    return ret;
408}
409
410int ClockRecoveryLoop::applyRateLimitedSlew() {
411    Mutex::Autolock lock(&lock_);
412
413    int ret = next_slew_change_timeout_.msecTillTimeout();
414    if (!ret) {
415        if (applySlew_l())
416            next_slew_change_timeout_.setTimeout(-1);
417        ret = next_slew_change_timeout_.msecTillTimeout();
418    }
419
420    return ret;
421}
422
423}  // namespace android
424