Patrick Ohly | a75244c | 2009-02-12 05:03:35 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (C) 2009 Intel Corporation. |
| 3 | * Author: Patrick Ohly <patrick.ohly@intel.com> |
| 4 | * |
| 5 | * This program is free software; you can redistribute it and/or modify |
| 6 | * it under the terms of the GNU General Public License as published by |
| 7 | * the Free Software Foundation; either version 2 of the License, or |
| 8 | * (at your option) any later version. |
| 9 | * |
| 10 | * This program is distributed in the hope that it will be useful, |
| 11 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 12 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 13 | * GNU General Public License for more details. |
| 14 | * |
| 15 | * You should have received a copy of the GNU General Public License |
| 16 | * along with this program; if not, write to the Free Software |
| 17 | * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
| 18 | */ |
| 19 | |
| 20 | #include <linux/timecompare.h> |
| 21 | #include <linux/module.h> |
| 22 | #include <linux/math64.h> |
| 23 | |
| 24 | /* |
| 25 | * fixed point arithmetic scale factor for skew |
| 26 | * |
| 27 | * Usually one would measure skew in ppb (parts per billion, 1e9), but |
| 28 | * using a factor of 2 simplifies the math. |
| 29 | */ |
| 30 | #define TIMECOMPARE_SKEW_RESOLUTION (((s64)1)<<30) |
| 31 | |
| 32 | ktime_t timecompare_transform(struct timecompare *sync, |
| 33 | u64 source_tstamp) |
| 34 | { |
| 35 | u64 nsec; |
| 36 | |
| 37 | nsec = source_tstamp + sync->offset; |
| 38 | nsec += (s64)(source_tstamp - sync->last_update) * sync->skew / |
| 39 | TIMECOMPARE_SKEW_RESOLUTION; |
| 40 | |
| 41 | return ns_to_ktime(nsec); |
| 42 | } |
| 43 | EXPORT_SYMBOL(timecompare_transform); |
| 44 | |
| 45 | int timecompare_offset(struct timecompare *sync, |
| 46 | s64 *offset, |
| 47 | u64 *source_tstamp) |
| 48 | { |
| 49 | u64 start_source = 0, end_source = 0; |
| 50 | struct { |
| 51 | s64 offset; |
| 52 | s64 duration_target; |
| 53 | } buffer[10], sample, *samples; |
| 54 | int counter = 0, i; |
| 55 | int used; |
| 56 | int index; |
| 57 | int num_samples = sync->num_samples; |
| 58 | |
| 59 | if (num_samples > sizeof(buffer)/sizeof(buffer[0])) { |
| 60 | samples = kmalloc(sizeof(*samples) * num_samples, GFP_ATOMIC); |
| 61 | if (!samples) { |
| 62 | samples = buffer; |
| 63 | num_samples = sizeof(buffer)/sizeof(buffer[0]); |
| 64 | } |
| 65 | } else { |
| 66 | samples = buffer; |
| 67 | } |
| 68 | |
| 69 | /* run until we have enough valid samples, but do not try forever */ |
| 70 | i = 0; |
| 71 | counter = 0; |
| 72 | while (1) { |
| 73 | u64 ts; |
| 74 | ktime_t start, end; |
| 75 | |
| 76 | start = sync->target(); |
| 77 | ts = timecounter_read(sync->source); |
| 78 | end = sync->target(); |
| 79 | |
| 80 | if (!i) |
| 81 | start_source = ts; |
| 82 | |
| 83 | /* ignore negative durations */ |
| 84 | sample.duration_target = ktime_to_ns(ktime_sub(end, start)); |
| 85 | if (sample.duration_target >= 0) { |
| 86 | /* |
| 87 | * assume symetric delay to and from source: |
| 88 | * average target time corresponds to measured |
| 89 | * source time |
| 90 | */ |
| 91 | sample.offset = |
| 92 | ktime_to_ns(ktime_add(end, start)) / 2 - |
| 93 | ts; |
| 94 | |
| 95 | /* simple insertion sort based on duration */ |
| 96 | index = counter - 1; |
| 97 | while (index >= 0) { |
| 98 | if (samples[index].duration_target < |
| 99 | sample.duration_target) |
| 100 | break; |
| 101 | samples[index + 1] = samples[index]; |
| 102 | index--; |
| 103 | } |
| 104 | samples[index + 1] = sample; |
| 105 | counter++; |
| 106 | } |
| 107 | |
| 108 | i++; |
| 109 | if (counter >= num_samples || i >= 100000) { |
| 110 | end_source = ts; |
| 111 | break; |
| 112 | } |
| 113 | } |
| 114 | |
| 115 | *source_tstamp = (end_source + start_source) / 2; |
| 116 | |
| 117 | /* remove outliers by only using 75% of the samples */ |
| 118 | used = counter * 3 / 4; |
| 119 | if (!used) |
| 120 | used = counter; |
| 121 | if (used) { |
| 122 | /* calculate average */ |
| 123 | s64 off = 0; |
| 124 | for (index = 0; index < used; index++) |
| 125 | off += samples[index].offset; |
| 126 | *offset = div_s64(off, used); |
| 127 | } |
| 128 | |
| 129 | if (samples && samples != buffer) |
| 130 | kfree(samples); |
| 131 | |
| 132 | return used; |
| 133 | } |
| 134 | EXPORT_SYMBOL(timecompare_offset); |
| 135 | |
| 136 | void __timecompare_update(struct timecompare *sync, |
| 137 | u64 source_tstamp) |
| 138 | { |
| 139 | s64 offset; |
| 140 | u64 average_time; |
| 141 | |
| 142 | if (!timecompare_offset(sync, &offset, &average_time)) |
| 143 | return; |
| 144 | |
| 145 | if (!sync->last_update) { |
| 146 | sync->last_update = average_time; |
| 147 | sync->offset = offset; |
| 148 | sync->skew = 0; |
| 149 | } else { |
| 150 | s64 delta_nsec = average_time - sync->last_update; |
| 151 | |
| 152 | /* avoid division by negative or small deltas */ |
| 153 | if (delta_nsec >= 10000) { |
| 154 | s64 delta_offset_nsec = offset - sync->offset; |
| 155 | s64 skew; /* delta_offset_nsec * |
| 156 | TIMECOMPARE_SKEW_RESOLUTION / |
| 157 | delta_nsec */ |
| 158 | u64 divisor; |
| 159 | |
| 160 | /* div_s64() is limited to 32 bit divisor */ |
| 161 | skew = delta_offset_nsec * TIMECOMPARE_SKEW_RESOLUTION; |
| 162 | divisor = delta_nsec; |
| 163 | while (unlikely(divisor >= ((s64)1) << 32)) { |
| 164 | /* divide both by 2; beware, right shift |
| 165 | of negative value has undefined |
| 166 | behavior and can only be used for |
| 167 | the positive divisor */ |
| 168 | skew = div_s64(skew, 2); |
| 169 | divisor >>= 1; |
| 170 | } |
| 171 | skew = div_s64(skew, divisor); |
| 172 | |
| 173 | /* |
| 174 | * Calculate new overall skew as 4/16 the |
| 175 | * old value and 12/16 the new one. This is |
| 176 | * a rather arbitrary tradeoff between |
| 177 | * only using the latest measurement (0/16 and |
| 178 | * 16/16) and even more weight on past measurements. |
| 179 | */ |
| 180 | #define TIMECOMPARE_NEW_SKEW_PER_16 12 |
| 181 | sync->skew = |
| 182 | div_s64((16 - TIMECOMPARE_NEW_SKEW_PER_16) * |
| 183 | sync->skew + |
| 184 | TIMECOMPARE_NEW_SKEW_PER_16 * skew, |
| 185 | 16); |
| 186 | sync->last_update = average_time; |
| 187 | sync->offset = offset; |
| 188 | } |
| 189 | } |
| 190 | } |
| 191 | EXPORT_SYMBOL(__timecompare_update); |