xref: /openbmc/qemu/migration/dirtyrate.c (revision e6b5a071) 
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 
15 #include <zlib.h>
16 #include "qapi/error.h"
17 #include "cpu.h"
18 #include "qemu/config-file.h"
19 #include "exec/memory.h"
20 #include "exec/ramblock.h"
21 #include "exec/target_page.h"
22 #include "qemu/rcu_queue.h"
23 #include "qapi/qapi-commands-migration.h"
24 #include "migration.h"
25 #include "ram.h"
26 #include "trace.h"
27 #include "dirtyrate.h"
28 
29 static int CalculatingState = DIRTY_RATE_STATUS_UNSTARTED;
30 static struct DirtyRateStat DirtyStat;
31 
32 static int64_t set_sample_page_period(int64_t msec, int64_t initial_time)
33 {
34     int64_t current_time;
35 
36     current_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
37     if ((current_time - initial_time) >= msec) {
38         msec = current_time - initial_time;
39     } else {
40         g_usleep((msec + initial_time - current_time) * 1000);
41     }
42 
43     return msec;
44 }
45 
46 static bool is_sample_period_valid(int64_t sec)
47 {
48     if (sec < MIN_FETCH_DIRTYRATE_TIME_SEC ||
49         sec > MAX_FETCH_DIRTYRATE_TIME_SEC) {
50         return false;
51     }
52 
53     return true;
54 }
55 
56 static int dirtyrate_set_state(int *state, int old_state, int new_state)
57 {
58     assert(new_state < DIRTY_RATE_STATUS__MAX);
59     trace_dirtyrate_set_state(DirtyRateStatus_str(new_state));
60     if (qatomic_cmpxchg(state, old_state, new_state) == old_state) {
61         return 0;
62     } else {
63         return -1;
64     }
65 }
66 
67 static struct DirtyRateInfo *query_dirty_rate_info(void)
68 {
69     int64_t dirty_rate = DirtyStat.dirty_rate;
70     struct DirtyRateInfo *info = g_malloc0(sizeof(DirtyRateInfo));
71 
72     if (qatomic_read(&CalculatingState) == DIRTY_RATE_STATUS_MEASURED) {
73         info->has_dirty_rate = true;
74         info->dirty_rate = dirty_rate;
75     }
76 
77     info->status = CalculatingState;
78     info->start_time = DirtyStat.start_time;
79     info->calc_time = DirtyStat.calc_time;
80 
81     trace_query_dirty_rate_info(DirtyRateStatus_str(CalculatingState));
82 
83     return info;
84 }
85 
86 static void init_dirtyrate_stat(int64_t start_time, int64_t calc_time)
87 {
88     DirtyStat.total_dirty_samples = 0;
89     DirtyStat.total_sample_count = 0;
90     DirtyStat.total_block_mem_MB = 0;
91     DirtyStat.dirty_rate = -1;
92     DirtyStat.start_time = start_time;
93     DirtyStat.calc_time = calc_time;
94 }
95 
96 static void update_dirtyrate_stat(struct RamblockDirtyInfo *info)
97 {
98     DirtyStat.total_dirty_samples += info->sample_dirty_count;
99     DirtyStat.total_sample_count += info->sample_pages_count;
100     /* size of total pages in MB */
101     DirtyStat.total_block_mem_MB += (info->ramblock_pages *
102                                      TARGET_PAGE_SIZE) >> 20;
103 }
104 
105 static void update_dirtyrate(uint64_t msec)
106 {
107     uint64_t dirtyrate;
108     uint64_t total_dirty_samples = DirtyStat.total_dirty_samples;
109     uint64_t total_sample_count = DirtyStat.total_sample_count;
110     uint64_t total_block_mem_MB = DirtyStat.total_block_mem_MB;
111 
112     dirtyrate = total_dirty_samples * total_block_mem_MB *
113                 1000 / (total_sample_count * msec);
114 
115     DirtyStat.dirty_rate = dirtyrate;
116 }
117 
118 /*
119  * get hash result for the sampled memory with length of TARGET_PAGE_SIZE
120  * in ramblock, which starts from ramblock base address.
121  */
122 static uint32_t get_ramblock_vfn_hash(struct RamblockDirtyInfo *info,
123                                       uint64_t vfn)
124 {
125     uint32_t crc;
126 
127     crc = crc32(0, (info->ramblock_addr +
128                 vfn * TARGET_PAGE_SIZE), TARGET_PAGE_SIZE);
129 
130     trace_get_ramblock_vfn_hash(info->idstr, vfn, crc);
131     return crc;
132 }
133 
134 static bool save_ramblock_hash(struct RamblockDirtyInfo *info)
135 {
136     unsigned int sample_pages_count;
137     int i;
138     GRand *rand;
139 
140     sample_pages_count = info->sample_pages_count;
141 
142     /* ramblock size less than one page, return success to skip this ramblock */
143     if (unlikely(info->ramblock_pages == 0 || sample_pages_count == 0)) {
144         return true;
145     }
146 
147     info->hash_result = g_try_malloc0_n(sample_pages_count,
148                                         sizeof(uint32_t));
149     if (!info->hash_result) {
150         return false;
151     }
152 
153     info->sample_page_vfn = g_try_malloc0_n(sample_pages_count,
154                                             sizeof(uint64_t));
155     if (!info->sample_page_vfn) {
156         g_free(info->hash_result);
157         return false;
158     }
159 
160     rand  = g_rand_new();
161     for (i = 0; i < sample_pages_count; i++) {
162         info->sample_page_vfn[i] = g_rand_int_range(rand, 0,
163                                                     info->ramblock_pages - 1);
164         info->hash_result[i] = get_ramblock_vfn_hash(info,
165                                                      info->sample_page_vfn[i]);
166     }
167     g_rand_free(rand);
168 
169     return true;
170 }
171 
172 static void get_ramblock_dirty_info(RAMBlock *block,
173                                     struct RamblockDirtyInfo *info,
174                                     struct DirtyRateConfig *config)
175 {
176     uint64_t sample_pages_per_gigabytes = config->sample_pages_per_gigabytes;
177 
178     /* Right shift 30 bits to calc ramblock size in GB */
179     info->sample_pages_count = (qemu_ram_get_used_length(block) *
180                                 sample_pages_per_gigabytes) >> 30;
181     /* Right shift TARGET_PAGE_BITS to calc page count */
182     info->ramblock_pages = qemu_ram_get_used_length(block) >>
183                            TARGET_PAGE_BITS;
184     info->ramblock_addr = qemu_ram_get_host_addr(block);
185     strcpy(info->idstr, qemu_ram_get_idstr(block));
186 }
187 
188 static void free_ramblock_dirty_info(struct RamblockDirtyInfo *infos, int count)
189 {
190     int i;
191 
192     if (!infos) {
193         return;
194     }
195 
196     for (i = 0; i < count; i++) {
197         g_free(infos[i].sample_page_vfn);
198         g_free(infos[i].hash_result);
199     }
200     g_free(infos);
201 }
202 
203 static bool skip_sample_ramblock(RAMBlock *block)
204 {
205     /*
206      * Sample only blocks larger than MIN_RAMBLOCK_SIZE.
207      */
208     if (qemu_ram_get_used_length(block) < (MIN_RAMBLOCK_SIZE << 10)) {
209         trace_skip_sample_ramblock(block->idstr,
210                                    qemu_ram_get_used_length(block));
211         return true;
212     }
213 
214     return false;
215 }
216 
217 static bool record_ramblock_hash_info(struct RamblockDirtyInfo **block_dinfo,
218                                       struct DirtyRateConfig config,
219                                       int *block_count)
220 {
221     struct RamblockDirtyInfo *info = NULL;
222     struct RamblockDirtyInfo *dinfo = NULL;
223     RAMBlock *block = NULL;
224     int total_count = 0;
225     int index = 0;
226     bool ret = false;
227 
228     RAMBLOCK_FOREACH_MIGRATABLE(block) {
229         if (skip_sample_ramblock(block)) {
230             continue;
231         }
232         total_count++;
233     }
234 
235     dinfo = g_try_malloc0_n(total_count, sizeof(struct RamblockDirtyInfo));
236     if (dinfo == NULL) {
237         goto out;
238     }
239 
240     RAMBLOCK_FOREACH_MIGRATABLE(block) {
241         if (skip_sample_ramblock(block)) {
242             continue;
243         }
244         if (index >= total_count) {
245             break;
246         }
247         info = &dinfo[index];
248         get_ramblock_dirty_info(block, info, &config);
249         if (!save_ramblock_hash(info)) {
250             goto out;
251         }
252         index++;
253     }
254     ret = true;
255 
256 out:
257     *block_count = index;
258     *block_dinfo = dinfo;
259     return ret;
260 }
261 
262 static void calc_page_dirty_rate(struct RamblockDirtyInfo *info)
263 {
264     uint32_t crc;
265     int i;
266 
267     for (i = 0; i < info->sample_pages_count; i++) {
268         crc = get_ramblock_vfn_hash(info, info->sample_page_vfn[i]);
269         if (crc != info->hash_result[i]) {
270             trace_calc_page_dirty_rate(info->idstr, crc, info->hash_result[i]);
271             info->sample_dirty_count++;
272         }
273     }
274 }
275 
276 static struct RamblockDirtyInfo *
277 find_block_matched(RAMBlock *block, int count,
278                   struct RamblockDirtyInfo *infos)
279 {
280     int i;
281     struct RamblockDirtyInfo *matched;
282 
283     for (i = 0; i < count; i++) {
284         if (!strcmp(infos[i].idstr, qemu_ram_get_idstr(block))) {
285             break;
286         }
287     }
288 
289     if (i == count) {
290         return NULL;
291     }
292 
293     if (infos[i].ramblock_addr != qemu_ram_get_host_addr(block) ||
294         infos[i].ramblock_pages !=
295             (qemu_ram_get_used_length(block) >> TARGET_PAGE_BITS)) {
296         trace_find_page_matched(block->idstr);
297         return NULL;
298     }
299 
300     matched = &infos[i];
301 
302     return matched;
303 }
304 
305 static bool compare_page_hash_info(struct RamblockDirtyInfo *info,
306                                   int block_count)
307 {
308     struct RamblockDirtyInfo *block_dinfo = NULL;
309     RAMBlock *block = NULL;
310 
311     RAMBLOCK_FOREACH_MIGRATABLE(block) {
312         if (skip_sample_ramblock(block)) {
313             continue;
314         }
315         block_dinfo = find_block_matched(block, block_count, info);
316         if (block_dinfo == NULL) {
317             continue;
318         }
319         calc_page_dirty_rate(block_dinfo);
320         update_dirtyrate_stat(block_dinfo);
321     }
322 
323     if (DirtyStat.total_sample_count == 0) {
324         return false;
325     }
326 
327     return true;
328 }
329 
330 static void calculate_dirtyrate(struct DirtyRateConfig config)
331 {
332     struct RamblockDirtyInfo *block_dinfo = NULL;
333     int block_count = 0;
334     int64_t msec = 0;
335     int64_t initial_time;
336 
337     rcu_register_thread();
338     rcu_read_lock();
339     initial_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
340     if (!record_ramblock_hash_info(&block_dinfo, config, &block_count)) {
341         goto out;
342     }
343     rcu_read_unlock();
344 
345     msec = config.sample_period_seconds * 1000;
346     msec = set_sample_page_period(msec, initial_time);
347     DirtyStat.start_time = initial_time / 1000;
348     DirtyStat.calc_time = msec / 1000;
349 
350     rcu_read_lock();
351     if (!compare_page_hash_info(block_dinfo, block_count)) {
352         goto out;
353     }
354 
355     update_dirtyrate(msec);
356 
357 out:
358     rcu_read_unlock();
359     free_ramblock_dirty_info(block_dinfo, block_count);
360     rcu_unregister_thread();
361 }
362 
363 void *get_dirtyrate_thread(void *arg)
364 {
365     struct DirtyRateConfig config = *(struct DirtyRateConfig *)arg;
366     int ret;
367     int64_t start_time;
368     int64_t calc_time;
369 
370     ret = dirtyrate_set_state(&CalculatingState, DIRTY_RATE_STATUS_UNSTARTED,
371                               DIRTY_RATE_STATUS_MEASURING);
372     if (ret == -1) {
373         error_report("change dirtyrate state failed.");
374         return NULL;
375     }
376 
377     start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) / 1000;
378     calc_time = config.sample_period_seconds;
379     init_dirtyrate_stat(start_time, calc_time);
380 
381     calculate_dirtyrate(config);
382 
383     ret = dirtyrate_set_state(&CalculatingState, DIRTY_RATE_STATUS_MEASURING,
384                               DIRTY_RATE_STATUS_MEASURED);
385     if (ret == -1) {
386         error_report("change dirtyrate state failed.");
387     }
388     return NULL;
389 }
390 
391 void qmp_calc_dirty_rate(int64_t calc_time, Error **errp)
392 {
393     static struct DirtyRateConfig config;
394     QemuThread thread;
395     int ret;
396 
397     /*
398      * If the dirty rate is already being measured, don't attempt to start.
399      */
400     if (qatomic_read(&CalculatingState) == DIRTY_RATE_STATUS_MEASURING) {
401         error_setg(errp, "the dirty rate is already being measured.");
402         return;
403     }
404 
405     if (!is_sample_period_valid(calc_time)) {
406         error_setg(errp, "calc-time is out of range[%d, %d].",
407                          MIN_FETCH_DIRTYRATE_TIME_SEC,
408                          MAX_FETCH_DIRTYRATE_TIME_SEC);
409         return;
410     }
411 
412     /*
413      * Init calculation state as unstarted.
414      */
415     ret = dirtyrate_set_state(&CalculatingState, CalculatingState,
416                               DIRTY_RATE_STATUS_UNSTARTED);
417     if (ret == -1) {
418         error_setg(errp, "init dirty rate calculation state failed.");
419         return;
420     }
421 
422     config.sample_period_seconds = calc_time;
423     config.sample_pages_per_gigabytes = DIRTYRATE_DEFAULT_SAMPLE_PAGES;
424     qemu_thread_create(&thread, "get_dirtyrate", get_dirtyrate_thread,
425                        (void *)&config, QEMU_THREAD_DETACHED);
426 }
427 
428 struct DirtyRateInfo *qmp_query_dirty_rate(Error **errp)
429 {
430     return query_dirty_rate_info();
431 }
432