xref: /openbmc/qemu/migration/dirtyrate.c (revision fe1127da)
1 /*
2  * Dirtyrate implement code
3  *
4  * Copyright (c) 2020 HUAWEI TECHNOLOGIES CO.,LTD.
5  *
6  * Authors:
7  *  Chuan Zheng <zhengchuan@huawei.com>
8  *
9  * This work is licensed under the terms of the GNU GPL, version 2 or later.
10  * See the COPYING file in the top-level directory.
11  */
12 
13 #include "qemu/osdep.h"
14 #include <zlib.h>
15 #include "qapi/error.h"
16 #include "cpu.h"
17 #include "exec/ramblock.h"
18 #include "qemu/rcu_queue.h"
19 #include "qapi/qapi-commands-migration.h"
20 #include "ram.h"
21 #include "trace.h"
22 #include "dirtyrate.h"
23 
24 static int CalculatingState = DIRTY_RATE_STATUS_UNSTARTED;
25 static struct DirtyRateStat DirtyStat;
26 
27 static int64_t set_sample_page_period(int64_t msec, int64_t initial_time)
28 {
29     int64_t current_time;
30 
31     current_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
32     if ((current_time - initial_time) >= msec) {
33         msec = current_time - initial_time;
34     } else {
35         g_usleep((msec + initial_time - current_time) * 1000);
36     }
37 
38     return msec;
39 }
40 
41 static bool is_sample_period_valid(int64_t sec)
42 {
43     if (sec < MIN_FETCH_DIRTYRATE_TIME_SEC ||
44         sec > MAX_FETCH_DIRTYRATE_TIME_SEC) {
45         return false;
46     }
47 
48     return true;
49 }
50 
51 static int dirtyrate_set_state(int *state, int old_state, int new_state)
52 {
53     assert(new_state < DIRTY_RATE_STATUS__MAX);
54     trace_dirtyrate_set_state(DirtyRateStatus_str(new_state));
55     if (qatomic_cmpxchg(state, old_state, new_state) == old_state) {
56         return 0;
57     } else {
58         return -1;
59     }
60 }
61 
62 static struct DirtyRateInfo *query_dirty_rate_info(void)
63 {
64     int64_t dirty_rate = DirtyStat.dirty_rate;
65     struct DirtyRateInfo *info = g_malloc0(sizeof(DirtyRateInfo));
66 
67     if (qatomic_read(&CalculatingState) == DIRTY_RATE_STATUS_MEASURED) {
68         info->has_dirty_rate = true;
69         info->dirty_rate = dirty_rate;
70     }
71 
72     info->status = CalculatingState;
73     info->start_time = DirtyStat.start_time;
74     info->calc_time = DirtyStat.calc_time;
75 
76     trace_query_dirty_rate_info(DirtyRateStatus_str(CalculatingState));
77 
78     return info;
79 }
80 
81 static void init_dirtyrate_stat(int64_t start_time, int64_t calc_time)
82 {
83     DirtyStat.total_dirty_samples = 0;
84     DirtyStat.total_sample_count = 0;
85     DirtyStat.total_block_mem_MB = 0;
86     DirtyStat.dirty_rate = -1;
87     DirtyStat.start_time = start_time;
88     DirtyStat.calc_time = calc_time;
89 }
90 
91 static void update_dirtyrate_stat(struct RamblockDirtyInfo *info)
92 {
93     DirtyStat.total_dirty_samples += info->sample_dirty_count;
94     DirtyStat.total_sample_count += info->sample_pages_count;
95     /* size of total pages in MB */
96     DirtyStat.total_block_mem_MB += (info->ramblock_pages *
97                                      TARGET_PAGE_SIZE) >> 20;
98 }
99 
100 static void update_dirtyrate(uint64_t msec)
101 {
102     uint64_t dirtyrate;
103     uint64_t total_dirty_samples = DirtyStat.total_dirty_samples;
104     uint64_t total_sample_count = DirtyStat.total_sample_count;
105     uint64_t total_block_mem_MB = DirtyStat.total_block_mem_MB;
106 
107     dirtyrate = total_dirty_samples * total_block_mem_MB *
108                 1000 / (total_sample_count * msec);
109 
110     DirtyStat.dirty_rate = dirtyrate;
111 }
112 
113 /*
114  * get hash result for the sampled memory with length of TARGET_PAGE_SIZE
115  * in ramblock, which starts from ramblock base address.
116  */
117 static uint32_t get_ramblock_vfn_hash(struct RamblockDirtyInfo *info,
118                                       uint64_t vfn)
119 {
120     uint32_t crc;
121 
122     crc = crc32(0, (info->ramblock_addr +
123                 vfn * TARGET_PAGE_SIZE), TARGET_PAGE_SIZE);
124 
125     trace_get_ramblock_vfn_hash(info->idstr, vfn, crc);
126     return crc;
127 }
128 
129 static bool save_ramblock_hash(struct RamblockDirtyInfo *info)
130 {
131     unsigned int sample_pages_count;
132     int i;
133     GRand *rand;
134 
135     sample_pages_count = info->sample_pages_count;
136 
137     /* ramblock size less than one page, return success to skip this ramblock */
138     if (unlikely(info->ramblock_pages == 0 || sample_pages_count == 0)) {
139         return true;
140     }
141 
142     info->hash_result = g_try_malloc0_n(sample_pages_count,
143                                         sizeof(uint32_t));
144     if (!info->hash_result) {
145         return false;
146     }
147 
148     info->sample_page_vfn = g_try_malloc0_n(sample_pages_count,
149                                             sizeof(uint64_t));
150     if (!info->sample_page_vfn) {
151         g_free(info->hash_result);
152         return false;
153     }
154 
155     rand  = g_rand_new();
156     for (i = 0; i < sample_pages_count; i++) {
157         info->sample_page_vfn[i] = g_rand_int_range(rand, 0,
158                                                     info->ramblock_pages - 1);
159         info->hash_result[i] = get_ramblock_vfn_hash(info,
160                                                      info->sample_page_vfn[i]);
161     }
162     g_rand_free(rand);
163 
164     return true;
165 }
166 
167 static void get_ramblock_dirty_info(RAMBlock *block,
168                                     struct RamblockDirtyInfo *info,
169                                     struct DirtyRateConfig *config)
170 {
171     uint64_t sample_pages_per_gigabytes = config->sample_pages_per_gigabytes;
172 
173     /* Right shift 30 bits to calc ramblock size in GB */
174     info->sample_pages_count = (qemu_ram_get_used_length(block) *
175                                 sample_pages_per_gigabytes) >> 30;
176     /* Right shift TARGET_PAGE_BITS to calc page count */
177     info->ramblock_pages = qemu_ram_get_used_length(block) >>
178                            TARGET_PAGE_BITS;
179     info->ramblock_addr = qemu_ram_get_host_addr(block);
180     strcpy(info->idstr, qemu_ram_get_idstr(block));
181 }
182 
183 static void free_ramblock_dirty_info(struct RamblockDirtyInfo *infos, int count)
184 {
185     int i;
186 
187     if (!infos) {
188         return;
189     }
190 
191     for (i = 0; i < count; i++) {
192         g_free(infos[i].sample_page_vfn);
193         g_free(infos[i].hash_result);
194     }
195     g_free(infos);
196 }
197 
198 static bool skip_sample_ramblock(RAMBlock *block)
199 {
200     /*
201      * Sample only blocks larger than MIN_RAMBLOCK_SIZE.
202      */
203     if (qemu_ram_get_used_length(block) < (MIN_RAMBLOCK_SIZE << 10)) {
204         trace_skip_sample_ramblock(block->idstr,
205                                    qemu_ram_get_used_length(block));
206         return true;
207     }
208 
209     return false;
210 }
211 
212 static bool record_ramblock_hash_info(struct RamblockDirtyInfo **block_dinfo,
213                                       struct DirtyRateConfig config,
214                                       int *block_count)
215 {
216     struct RamblockDirtyInfo *info = NULL;
217     struct RamblockDirtyInfo *dinfo = NULL;
218     RAMBlock *block = NULL;
219     int total_count = 0;
220     int index = 0;
221     bool ret = false;
222 
223     RAMBLOCK_FOREACH_MIGRATABLE(block) {
224         if (skip_sample_ramblock(block)) {
225             continue;
226         }
227         total_count++;
228     }
229 
230     dinfo = g_try_malloc0_n(total_count, sizeof(struct RamblockDirtyInfo));
231     if (dinfo == NULL) {
232         goto out;
233     }
234 
235     RAMBLOCK_FOREACH_MIGRATABLE(block) {
236         if (skip_sample_ramblock(block)) {
237             continue;
238         }
239         if (index >= total_count) {
240             break;
241         }
242         info = &dinfo[index];
243         get_ramblock_dirty_info(block, info, &config);
244         if (!save_ramblock_hash(info)) {
245             goto out;
246         }
247         index++;
248     }
249     ret = true;
250 
251 out:
252     *block_count = index;
253     *block_dinfo = dinfo;
254     return ret;
255 }
256 
257 static void calc_page_dirty_rate(struct RamblockDirtyInfo *info)
258 {
259     uint32_t crc;
260     int i;
261 
262     for (i = 0; i < info->sample_pages_count; i++) {
263         crc = get_ramblock_vfn_hash(info, info->sample_page_vfn[i]);
264         if (crc != info->hash_result[i]) {
265             trace_calc_page_dirty_rate(info->idstr, crc, info->hash_result[i]);
266             info->sample_dirty_count++;
267         }
268     }
269 }
270 
271 static struct RamblockDirtyInfo *
272 find_block_matched(RAMBlock *block, int count,
273                   struct RamblockDirtyInfo *infos)
274 {
275     int i;
276     struct RamblockDirtyInfo *matched;
277 
278     for (i = 0; i < count; i++) {
279         if (!strcmp(infos[i].idstr, qemu_ram_get_idstr(block))) {
280             break;
281         }
282     }
283 
284     if (i == count) {
285         return NULL;
286     }
287 
288     if (infos[i].ramblock_addr != qemu_ram_get_host_addr(block) ||
289         infos[i].ramblock_pages !=
290             (qemu_ram_get_used_length(block) >> TARGET_PAGE_BITS)) {
291         trace_find_page_matched(block->idstr);
292         return NULL;
293     }
294 
295     matched = &infos[i];
296 
297     return matched;
298 }
299 
300 static bool compare_page_hash_info(struct RamblockDirtyInfo *info,
301                                   int block_count)
302 {
303     struct RamblockDirtyInfo *block_dinfo = NULL;
304     RAMBlock *block = NULL;
305 
306     RAMBLOCK_FOREACH_MIGRATABLE(block) {
307         if (skip_sample_ramblock(block)) {
308             continue;
309         }
310         block_dinfo = find_block_matched(block, block_count, info);
311         if (block_dinfo == NULL) {
312             continue;
313         }
314         calc_page_dirty_rate(block_dinfo);
315         update_dirtyrate_stat(block_dinfo);
316     }
317 
318     if (DirtyStat.total_sample_count == 0) {
319         return false;
320     }
321 
322     return true;
323 }
324 
325 static void calculate_dirtyrate(struct DirtyRateConfig config)
326 {
327     struct RamblockDirtyInfo *block_dinfo = NULL;
328     int block_count = 0;
329     int64_t msec = 0;
330     int64_t initial_time;
331 
332     rcu_register_thread();
333     rcu_read_lock();
334     initial_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
335     if (!record_ramblock_hash_info(&block_dinfo, config, &block_count)) {
336         goto out;
337     }
338     rcu_read_unlock();
339 
340     msec = config.sample_period_seconds * 1000;
341     msec = set_sample_page_period(msec, initial_time);
342     DirtyStat.start_time = initial_time / 1000;
343     DirtyStat.calc_time = msec / 1000;
344 
345     rcu_read_lock();
346     if (!compare_page_hash_info(block_dinfo, block_count)) {
347         goto out;
348     }
349 
350     update_dirtyrate(msec);
351 
352 out:
353     rcu_read_unlock();
354     free_ramblock_dirty_info(block_dinfo, block_count);
355     rcu_unregister_thread();
356 }
357 
358 void *get_dirtyrate_thread(void *arg)
359 {
360     struct DirtyRateConfig config = *(struct DirtyRateConfig *)arg;
361     int ret;
362     int64_t start_time;
363     int64_t calc_time;
364 
365     ret = dirtyrate_set_state(&CalculatingState, DIRTY_RATE_STATUS_UNSTARTED,
366                               DIRTY_RATE_STATUS_MEASURING);
367     if (ret == -1) {
368         error_report("change dirtyrate state failed.");
369         return NULL;
370     }
371 
372     start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) / 1000;
373     calc_time = config.sample_period_seconds;
374     init_dirtyrate_stat(start_time, calc_time);
375 
376     calculate_dirtyrate(config);
377 
378     ret = dirtyrate_set_state(&CalculatingState, DIRTY_RATE_STATUS_MEASURING,
379                               DIRTY_RATE_STATUS_MEASURED);
380     if (ret == -1) {
381         error_report("change dirtyrate state failed.");
382     }
383     return NULL;
384 }
385 
386 void qmp_calc_dirty_rate(int64_t calc_time, Error **errp)
387 {
388     static struct DirtyRateConfig config;
389     QemuThread thread;
390     int ret;
391 
392     /*
393      * If the dirty rate is already being measured, don't attempt to start.
394      */
395     if (qatomic_read(&CalculatingState) == DIRTY_RATE_STATUS_MEASURING) {
396         error_setg(errp, "the dirty rate is already being measured.");
397         return;
398     }
399 
400     if (!is_sample_period_valid(calc_time)) {
401         error_setg(errp, "calc-time is out of range[%d, %d].",
402                          MIN_FETCH_DIRTYRATE_TIME_SEC,
403                          MAX_FETCH_DIRTYRATE_TIME_SEC);
404         return;
405     }
406 
407     /*
408      * Init calculation state as unstarted.
409      */
410     ret = dirtyrate_set_state(&CalculatingState, CalculatingState,
411                               DIRTY_RATE_STATUS_UNSTARTED);
412     if (ret == -1) {
413         error_setg(errp, "init dirty rate calculation state failed.");
414         return;
415     }
416 
417     config.sample_period_seconds = calc_time;
418     config.sample_pages_per_gigabytes = DIRTYRATE_DEFAULT_SAMPLE_PAGES;
419     qemu_thread_create(&thread, "get_dirtyrate", get_dirtyrate_thread,
420                        (void *)&config, QEMU_THREAD_DETACHED);
421 }
422 
423 struct DirtyRateInfo *qmp_query_dirty_rate(Error **errp)
424 {
425     return query_dirty_rate_info();
426 }
427