xref: /openbmc/qemu/block/backup.c (revision 1b111dc1)
1 /*
2  * QEMU backup
3  *
4  * Copyright (C) 2013 Proxmox Server Solutions
5  *
6  * Authors:
7  *  Dietmar Maurer (dietmar@proxmox.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 
14 #include <stdio.h>
15 #include <errno.h>
16 #include <unistd.h>
17 
18 #include "trace.h"
19 #include "block/block.h"
20 #include "block/block_int.h"
21 #include "block/blockjob.h"
22 #include "qemu/ratelimit.h"
23 
24 #define BACKUP_CLUSTER_BITS 16
25 #define BACKUP_CLUSTER_SIZE (1 << BACKUP_CLUSTER_BITS)
26 #define BACKUP_SECTORS_PER_CLUSTER (BACKUP_CLUSTER_SIZE / BDRV_SECTOR_SIZE)
27 
28 #define SLICE_TIME 100000000ULL /* ns */
29 
30 typedef struct CowRequest {
31     int64_t start;
32     int64_t end;
33     QLIST_ENTRY(CowRequest) list;
34     CoQueue wait_queue; /* coroutines blocked on this request */
35 } CowRequest;
36 
37 typedef struct BackupBlockJob {
38     BlockJob common;
39     BlockDriverState *target;
40     MirrorSyncMode sync_mode;
41     RateLimit limit;
42     BlockdevOnError on_source_error;
43     BlockdevOnError on_target_error;
44     CoRwlock flush_rwlock;
45     uint64_t sectors_read;
46     HBitmap *bitmap;
47     QLIST_HEAD(, CowRequest) inflight_reqs;
48 } BackupBlockJob;
49 
50 /* See if in-flight requests overlap and wait for them to complete */
51 static void coroutine_fn wait_for_overlapping_requests(BackupBlockJob *job,
52                                                        int64_t start,
53                                                        int64_t end)
54 {
55     CowRequest *req;
56     bool retry;
57 
58     do {
59         retry = false;
60         QLIST_FOREACH(req, &job->inflight_reqs, list) {
61             if (end > req->start && start < req->end) {
62                 qemu_co_queue_wait(&req->wait_queue);
63                 retry = true;
64                 break;
65             }
66         }
67     } while (retry);
68 }
69 
70 /* Keep track of an in-flight request */
71 static void cow_request_begin(CowRequest *req, BackupBlockJob *job,
72                                      int64_t start, int64_t end)
73 {
74     req->start = start;
75     req->end = end;
76     qemu_co_queue_init(&req->wait_queue);
77     QLIST_INSERT_HEAD(&job->inflight_reqs, req, list);
78 }
79 
80 /* Forget about a completed request */
81 static void cow_request_end(CowRequest *req)
82 {
83     QLIST_REMOVE(req, list);
84     qemu_co_queue_restart_all(&req->wait_queue);
85 }
86 
87 static int coroutine_fn backup_do_cow(BlockDriverState *bs,
88                                       int64_t sector_num, int nb_sectors,
89                                       bool *error_is_read)
90 {
91     BackupBlockJob *job = (BackupBlockJob *)bs->job;
92     CowRequest cow_request;
93     struct iovec iov;
94     QEMUIOVector bounce_qiov;
95     void *bounce_buffer = NULL;
96     int ret = 0;
97     int64_t start, end;
98     int n;
99 
100     qemu_co_rwlock_rdlock(&job->flush_rwlock);
101 
102     start = sector_num / BACKUP_SECTORS_PER_CLUSTER;
103     end = DIV_ROUND_UP(sector_num + nb_sectors, BACKUP_SECTORS_PER_CLUSTER);
104 
105     trace_backup_do_cow_enter(job, start, sector_num, nb_sectors);
106 
107     wait_for_overlapping_requests(job, start, end);
108     cow_request_begin(&cow_request, job, start, end);
109 
110     for (; start < end; start++) {
111         if (hbitmap_get(job->bitmap, start)) {
112             trace_backup_do_cow_skip(job, start);
113             continue; /* already copied */
114         }
115 
116         trace_backup_do_cow_process(job, start);
117 
118         n = MIN(BACKUP_SECTORS_PER_CLUSTER,
119                 job->common.len / BDRV_SECTOR_SIZE -
120                 start * BACKUP_SECTORS_PER_CLUSTER);
121 
122         if (!bounce_buffer) {
123             bounce_buffer = qemu_blockalign(bs, BACKUP_CLUSTER_SIZE);
124         }
125         iov.iov_base = bounce_buffer;
126         iov.iov_len = n * BDRV_SECTOR_SIZE;
127         qemu_iovec_init_external(&bounce_qiov, &iov, 1);
128 
129         ret = bdrv_co_readv(bs, start * BACKUP_SECTORS_PER_CLUSTER, n,
130                             &bounce_qiov);
131         if (ret < 0) {
132             trace_backup_do_cow_read_fail(job, start, ret);
133             if (error_is_read) {
134                 *error_is_read = true;
135             }
136             goto out;
137         }
138 
139         if (buffer_is_zero(iov.iov_base, iov.iov_len)) {
140             ret = bdrv_co_write_zeroes(job->target,
141                                        start * BACKUP_SECTORS_PER_CLUSTER,
142                                        n, BDRV_REQ_MAY_UNMAP);
143         } else {
144             ret = bdrv_co_writev(job->target,
145                                  start * BACKUP_SECTORS_PER_CLUSTER, n,
146                                  &bounce_qiov);
147         }
148         if (ret < 0) {
149             trace_backup_do_cow_write_fail(job, start, ret);
150             if (error_is_read) {
151                 *error_is_read = false;
152             }
153             goto out;
154         }
155 
156         hbitmap_set(job->bitmap, start, 1);
157 
158         /* Publish progress, guest I/O counts as progress too.  Note that the
159          * offset field is an opaque progress value, it is not a disk offset.
160          */
161         job->sectors_read += n;
162         job->common.offset += n * BDRV_SECTOR_SIZE;
163     }
164 
165 out:
166     if (bounce_buffer) {
167         qemu_vfree(bounce_buffer);
168     }
169 
170     cow_request_end(&cow_request);
171 
172     trace_backup_do_cow_return(job, sector_num, nb_sectors, ret);
173 
174     qemu_co_rwlock_unlock(&job->flush_rwlock);
175 
176     return ret;
177 }
178 
179 static int coroutine_fn backup_before_write_notify(
180         NotifierWithReturn *notifier,
181         void *opaque)
182 {
183     BdrvTrackedRequest *req = opaque;
184 
185     return backup_do_cow(req->bs, req->sector_num, req->nb_sectors, NULL);
186 }
187 
188 static void backup_set_speed(BlockJob *job, int64_t speed, Error **errp)
189 {
190     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
191 
192     if (speed < 0) {
193         error_set(errp, QERR_INVALID_PARAMETER, "speed");
194         return;
195     }
196     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
197 }
198 
199 static void backup_iostatus_reset(BlockJob *job)
200 {
201     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
202 
203     bdrv_iostatus_reset(s->target);
204 }
205 
206 static const BlockJobDriver backup_job_driver = {
207     .instance_size  = sizeof(BackupBlockJob),
208     .job_type       = BLOCK_JOB_TYPE_BACKUP,
209     .set_speed      = backup_set_speed,
210     .iostatus_reset = backup_iostatus_reset,
211 };
212 
213 static BlockErrorAction backup_error_action(BackupBlockJob *job,
214                                             bool read, int error)
215 {
216     if (read) {
217         return block_job_error_action(&job->common, job->common.bs,
218                                       job->on_source_error, true, error);
219     } else {
220         return block_job_error_action(&job->common, job->target,
221                                       job->on_target_error, false, error);
222     }
223 }
224 
225 static void coroutine_fn backup_run(void *opaque)
226 {
227     BackupBlockJob *job = opaque;
228     BlockDriverState *bs = job->common.bs;
229     BlockDriverState *target = job->target;
230     BlockdevOnError on_target_error = job->on_target_error;
231     NotifierWithReturn before_write = {
232         .notify = backup_before_write_notify,
233     };
234     int64_t start, end;
235     int ret = 0;
236 
237     QLIST_INIT(&job->inflight_reqs);
238     qemu_co_rwlock_init(&job->flush_rwlock);
239 
240     start = 0;
241     end = DIV_ROUND_UP(job->common.len / BDRV_SECTOR_SIZE,
242                        BACKUP_SECTORS_PER_CLUSTER);
243 
244     job->bitmap = hbitmap_alloc(end, 0);
245 
246     bdrv_set_enable_write_cache(target, true);
247     bdrv_set_on_error(target, on_target_error, on_target_error);
248     bdrv_iostatus_enable(target);
249 
250     bdrv_add_before_write_notifier(bs, &before_write);
251 
252     if (job->sync_mode == MIRROR_SYNC_MODE_NONE) {
253         while (!block_job_is_cancelled(&job->common)) {
254             /* Yield until the job is cancelled.  We just let our before_write
255              * notify callback service CoW requests. */
256             job->common.busy = false;
257             qemu_coroutine_yield();
258             job->common.busy = true;
259         }
260     } else {
261         /* Both FULL and TOP SYNC_MODE's require copying.. */
262         for (; start < end; start++) {
263             bool error_is_read;
264 
265             if (block_job_is_cancelled(&job->common)) {
266                 break;
267             }
268 
269             /* we need to yield so that qemu_aio_flush() returns.
270              * (without, VM does not reboot)
271              */
272             if (job->common.speed) {
273                 uint64_t delay_ns = ratelimit_calculate_delay(
274                         &job->limit, job->sectors_read);
275                 job->sectors_read = 0;
276                 block_job_sleep_ns(&job->common, QEMU_CLOCK_REALTIME, delay_ns);
277             } else {
278                 block_job_sleep_ns(&job->common, QEMU_CLOCK_REALTIME, 0);
279             }
280 
281             if (block_job_is_cancelled(&job->common)) {
282                 break;
283             }
284 
285             if (job->sync_mode == MIRROR_SYNC_MODE_TOP) {
286                 int i, n;
287                 int alloced = 0;
288 
289                 /* Check to see if these blocks are already in the
290                  * backing file. */
291 
292                 for (i = 0; i < BACKUP_SECTORS_PER_CLUSTER;) {
293                     /* bdrv_is_allocated() only returns true/false based
294                      * on the first set of sectors it comes across that
295                      * are are all in the same state.
296                      * For that reason we must verify each sector in the
297                      * backup cluster length.  We end up copying more than
298                      * needed but at some point that is always the case. */
299                     alloced =
300                         bdrv_is_allocated(bs,
301                                 start * BACKUP_SECTORS_PER_CLUSTER + i,
302                                 BACKUP_SECTORS_PER_CLUSTER - i, &n);
303                     i += n;
304 
305                     if (alloced == 1) {
306                         break;
307                     }
308                 }
309 
310                 /* If the above loop never found any sectors that are in
311                  * the topmost image, skip this backup. */
312                 if (alloced == 0) {
313                     continue;
314                 }
315             }
316             /* FULL sync mode we copy the whole drive. */
317             ret = backup_do_cow(bs, start * BACKUP_SECTORS_PER_CLUSTER,
318                     BACKUP_SECTORS_PER_CLUSTER, &error_is_read);
319             if (ret < 0) {
320                 /* Depending on error action, fail now or retry cluster */
321                 BlockErrorAction action =
322                     backup_error_action(job, error_is_read, -ret);
323                 if (action == BDRV_ACTION_REPORT) {
324                     break;
325                 } else {
326                     start--;
327                     continue;
328                 }
329             }
330         }
331     }
332 
333     notifier_with_return_remove(&before_write);
334 
335     /* wait until pending backup_do_cow() calls have completed */
336     qemu_co_rwlock_wrlock(&job->flush_rwlock);
337     qemu_co_rwlock_unlock(&job->flush_rwlock);
338 
339     hbitmap_free(job->bitmap);
340 
341     bdrv_iostatus_disable(target);
342     bdrv_unref(target);
343 
344     block_job_completed(&job->common, ret);
345 }
346 
347 void backup_start(BlockDriverState *bs, BlockDriverState *target,
348                   int64_t speed, MirrorSyncMode sync_mode,
349                   BlockdevOnError on_source_error,
350                   BlockdevOnError on_target_error,
351                   BlockDriverCompletionFunc *cb, void *opaque,
352                   Error **errp)
353 {
354     int64_t len;
355 
356     assert(bs);
357     assert(target);
358     assert(cb);
359 
360     if ((on_source_error == BLOCKDEV_ON_ERROR_STOP ||
361          on_source_error == BLOCKDEV_ON_ERROR_ENOSPC) &&
362         !bdrv_iostatus_is_enabled(bs)) {
363         error_set(errp, QERR_INVALID_PARAMETER, "on-source-error");
364         return;
365     }
366 
367     len = bdrv_getlength(bs);
368     if (len < 0) {
369         error_setg_errno(errp, -len, "unable to get length for '%s'",
370                          bdrv_get_device_name(bs));
371         return;
372     }
373 
374     BackupBlockJob *job = block_job_create(&backup_job_driver, bs, speed,
375                                            cb, opaque, errp);
376     if (!job) {
377         return;
378     }
379 
380     job->on_source_error = on_source_error;
381     job->on_target_error = on_target_error;
382     job->target = target;
383     job->sync_mode = sync_mode;
384     job->common.len = len;
385     job->common.co = qemu_coroutine_create(backup_run);
386     qemu_coroutine_enter(job->common.co, job);
387 }
388