blob: 25f98e493c7c14ee66b0afb9a2da200610c9f725 [file] [log] [blame]
Stefan Hajnoczi4f1043b2012-01-18 14:40:44 +00001/*
2 * Image streaming
3 *
4 * Copyright IBM, Corp. 2011
5 *
6 * Authors:
7 * Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
8 *
9 * This work is licensed under the terms of the GNU LGPL, version 2 or later.
10 * See the COPYING.LIB file in the top-level directory.
11 *
12 */
13
14#include "trace.h"
15#include "block_int.h"
16
17enum {
18 /*
19 * Size of data buffer for populating the image file. This should be large
20 * enough to process multiple clusters in a single call, so that populating
21 * contiguous regions of the image is efficient.
22 */
23 STREAM_BUFFER_SIZE = 512 * 1024, /* in bytes */
24};
25
Stefan Hajnoczi5094a6c2012-01-18 14:40:45 +000026#define SLICE_TIME 100000000ULL /* ns */
27
28typedef struct {
29 int64_t next_slice_time;
30 uint64_t slice_quota;
31 uint64_t dispatched;
32} RateLimit;
33
34static int64_t ratelimit_calculate_delay(RateLimit *limit, uint64_t n)
35{
36 int64_t delay_ns = 0;
37 int64_t now = qemu_get_clock_ns(rt_clock);
38
39 if (limit->next_slice_time < now) {
40 limit->next_slice_time = now + SLICE_TIME;
41 limit->dispatched = 0;
42 }
43 if (limit->dispatched + n > limit->slice_quota) {
44 delay_ns = limit->next_slice_time - now;
45 } else {
46 limit->dispatched += n;
47 }
48 return delay_ns;
49}
50
51static void ratelimit_set_speed(RateLimit *limit, uint64_t speed)
52{
53 limit->slice_quota = speed / (1000000000ULL / SLICE_TIME);
54}
55
Stefan Hajnoczi4f1043b2012-01-18 14:40:44 +000056typedef struct StreamBlockJob {
57 BlockJob common;
Stefan Hajnoczi5094a6c2012-01-18 14:40:45 +000058 RateLimit limit;
Stefan Hajnoczi4f1043b2012-01-18 14:40:44 +000059 BlockDriverState *base;
Marcelo Tosattic8c30802012-01-18 14:40:53 +000060 char backing_file_id[1024];
Stefan Hajnoczi4f1043b2012-01-18 14:40:44 +000061} StreamBlockJob;
62
63static int coroutine_fn stream_populate(BlockDriverState *bs,
64 int64_t sector_num, int nb_sectors,
65 void *buf)
66{
67 struct iovec iov = {
68 .iov_base = buf,
69 .iov_len = nb_sectors * BDRV_SECTOR_SIZE,
70 };
71 QEMUIOVector qiov;
72
73 qemu_iovec_init_external(&qiov, &iov, 1);
74
75 /* Copy-on-read the unallocated clusters */
76 return bdrv_co_copy_on_readv(bs, sector_num, nb_sectors, &qiov);
77}
78
Marcelo Tosatti5a67a102012-03-26 21:22:10 -030079static void close_unused_images(BlockDriverState *top, BlockDriverState *base,
80 const char *base_id)
81{
82 BlockDriverState *intermediate;
83 intermediate = top->backing_hd;
84
85 while (intermediate) {
86 BlockDriverState *unused;
87
88 /* reached base */
89 if (intermediate == base) {
90 break;
91 }
92
93 unused = intermediate;
94 intermediate = intermediate->backing_hd;
95 unused->backing_hd = NULL;
96 bdrv_delete(unused);
97 }
98 top->backing_hd = base;
Marcelo Tosatti5a67a102012-03-26 21:22:10 -030099}
100
Marcelo Tosattic8c30802012-01-18 14:40:53 +0000101/*
102 * Given an image chain: [BASE] -> [INTER1] -> [INTER2] -> [TOP]
103 *
104 * Return true if the given sector is allocated in top.
105 * Return false if the given sector is allocated in intermediate images.
106 * Return true otherwise.
107 *
108 * 'pnum' is set to the number of sectors (including and immediately following
109 * the specified sector) that are known to be in the same
110 * allocated/unallocated state.
111 *
112 */
113static int coroutine_fn is_allocated_base(BlockDriverState *top,
114 BlockDriverState *base,
115 int64_t sector_num,
116 int nb_sectors, int *pnum)
117{
118 BlockDriverState *intermediate;
119 int ret, n;
120
121 ret = bdrv_co_is_allocated(top, sector_num, nb_sectors, &n);
122 if (ret) {
123 *pnum = n;
124 return ret;
125 }
126
127 /*
128 * Is the unallocated chunk [sector_num, n] also
129 * unallocated between base and top?
130 */
131 intermediate = top->backing_hd;
132
133 while (intermediate) {
134 int pnum_inter;
135
136 /* reached base */
137 if (intermediate == base) {
138 *pnum = n;
139 return 1;
140 }
141 ret = bdrv_co_is_allocated(intermediate, sector_num, nb_sectors,
142 &pnum_inter);
143 if (ret < 0) {
144 return ret;
145 } else if (ret) {
146 *pnum = pnum_inter;
147 return 0;
148 }
149
150 /*
151 * [sector_num, nb_sectors] is unallocated on top but intermediate
152 * might have
153 *
154 * [sector_num+x, nr_sectors] allocated.
155 */
156 if (n > pnum_inter) {
157 n = pnum_inter;
158 }
159
160 intermediate = intermediate->backing_hd;
161 }
162
163 return 1;
164}
165
Stefan Hajnoczi4f1043b2012-01-18 14:40:44 +0000166static void coroutine_fn stream_run(void *opaque)
167{
168 StreamBlockJob *s = opaque;
169 BlockDriverState *bs = s->common.bs;
Marcelo Tosattic8c30802012-01-18 14:40:53 +0000170 BlockDriverState *base = s->base;
Stefan Hajnoczi4f1043b2012-01-18 14:40:44 +0000171 int64_t sector_num, end;
172 int ret = 0;
173 int n;
174 void *buf;
175
176 s->common.len = bdrv_getlength(bs);
177 if (s->common.len < 0) {
178 block_job_complete(&s->common, s->common.len);
179 return;
180 }
181
182 end = s->common.len >> BDRV_SECTOR_BITS;
183 buf = qemu_blockalign(bs, STREAM_BUFFER_SIZE);
184
185 /* Turn on copy-on-read for the whole block device so that guest read
186 * requests help us make progress. Only do this when copying the entire
187 * backing chain since the copy-on-read operation does not take base into
188 * account.
189 */
190 if (!base) {
191 bdrv_enable_copy_on_read(bs);
192 }
193
194 for (sector_num = 0; sector_num < end; sector_num += n) {
Paolo Bonzini4513eaf2012-05-08 16:51:45 +0200195 uint64_t delay_ns = 0;
196
197wait:
198 /* Note that even when no rate limit is applied we need to yield
199 * with no pending I/O here so that qemu_aio_flush() returns.
200 */
201 block_job_sleep_ns(&s->common, rt_clock, delay_ns);
Stefan Hajnoczi4f1043b2012-01-18 14:40:44 +0000202 if (block_job_is_cancelled(&s->common)) {
203 break;
204 }
205
Marcelo Tosattic8c30802012-01-18 14:40:53 +0000206 if (base) {
207 ret = is_allocated_base(bs, base, sector_num,
208 STREAM_BUFFER_SIZE / BDRV_SECTOR_SIZE, &n);
209 } else {
210 ret = bdrv_co_is_allocated(bs, sector_num,
211 STREAM_BUFFER_SIZE / BDRV_SECTOR_SIZE,
212 &n);
213 }
Stefan Hajnoczi4f1043b2012-01-18 14:40:44 +0000214 trace_stream_one_iteration(s, sector_num, n, ret);
215 if (ret == 0) {
Stefan Hajnoczi5094a6c2012-01-18 14:40:45 +0000216 if (s->common.speed) {
Paolo Bonzini4513eaf2012-05-08 16:51:45 +0200217 delay_ns = ratelimit_calculate_delay(&s->limit, n);
Stefan Hajnoczi5094a6c2012-01-18 14:40:45 +0000218 if (delay_ns > 0) {
Paolo Bonzini4513eaf2012-05-08 16:51:45 +0200219 goto wait;
Stefan Hajnoczi5094a6c2012-01-18 14:40:45 +0000220 }
221 }
Stefan Hajnoczi4f1043b2012-01-18 14:40:44 +0000222 ret = stream_populate(bs, sector_num, n, buf);
223 }
224 if (ret < 0) {
225 break;
226 }
Marcelo Tosattic8c30802012-01-18 14:40:53 +0000227 ret = 0;
Stefan Hajnoczi4f1043b2012-01-18 14:40:44 +0000228
229 /* Publish progress */
230 s->common.offset += n * BDRV_SECTOR_SIZE;
231 }
232
233 if (!base) {
234 bdrv_disable_copy_on_read(bs);
235 }
236
Paolo Bonzini3e914652012-03-30 13:17:11 +0200237 if (!block_job_is_cancelled(&s->common) && sector_num == end && ret == 0) {
Paolo Bonzinif6133de2012-05-08 16:51:55 +0200238 const char *base_id = NULL, *base_fmt = NULL;
Marcelo Tosattic8c30802012-01-18 14:40:53 +0000239 if (base) {
240 base_id = s->backing_file_id;
Paolo Bonzinif6133de2012-05-08 16:51:55 +0200241 if (base->drv) {
242 base_fmt = base->drv->format_name;
243 }
Marcelo Tosattic8c30802012-01-18 14:40:53 +0000244 }
Paolo Bonzinif6133de2012-05-08 16:51:55 +0200245 ret = bdrv_change_backing_file(bs, base_id, base_fmt);
Marcelo Tosatti5a67a102012-03-26 21:22:10 -0300246 close_unused_images(bs, base, base_id);
Stefan Hajnoczi4f1043b2012-01-18 14:40:44 +0000247 }
248
249 qemu_vfree(buf);
250 block_job_complete(&s->common, ret);
251}
252
Stefan Hajnoczi882ec7c2012-04-25 16:51:02 +0100253static void stream_set_speed(BlockJob *job, int64_t speed, Error **errp)
Stefan Hajnoczi5094a6c2012-01-18 14:40:45 +0000254{
255 StreamBlockJob *s = container_of(job, StreamBlockJob, common);
256
Stefan Hajnoczi882ec7c2012-04-25 16:51:02 +0100257 if (speed < 0) {
258 error_set(errp, QERR_INVALID_PARAMETER, "speed");
Stefan Hajnoczi9e6636c2012-04-25 16:51:01 +0100259 return;
Stefan Hajnoczi5094a6c2012-01-18 14:40:45 +0000260 }
Stefan Hajnoczi882ec7c2012-04-25 16:51:02 +0100261 ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE);
Stefan Hajnoczi5094a6c2012-01-18 14:40:45 +0000262}
263
Stefan Hajnoczi4f1043b2012-01-18 14:40:44 +0000264static BlockJobType stream_job_type = {
265 .instance_size = sizeof(StreamBlockJob),
266 .job_type = "stream",
Stefan Hajnoczi5094a6c2012-01-18 14:40:45 +0000267 .set_speed = stream_set_speed,
Stefan Hajnoczi4f1043b2012-01-18 14:40:44 +0000268};
269
Stefan Hajnoczifd7f8c62012-04-25 16:51:00 +0100270void stream_start(BlockDriverState *bs, BlockDriverState *base,
Stefan Hajnoczic83c66c2012-04-25 16:51:03 +0100271 const char *base_id, int64_t speed,
272 BlockDriverCompletionFunc *cb,
Stefan Hajnoczifd7f8c62012-04-25 16:51:00 +0100273 void *opaque, Error **errp)
Stefan Hajnoczi4f1043b2012-01-18 14:40:44 +0000274{
275 StreamBlockJob *s;
Stefan Hajnoczi4f1043b2012-01-18 14:40:44 +0000276
Stefan Hajnoczic83c66c2012-04-25 16:51:03 +0100277 s = block_job_create(&stream_job_type, bs, speed, cb, opaque, errp);
Stefan Hajnoczi4f1043b2012-01-18 14:40:44 +0000278 if (!s) {
Stefan Hajnoczifd7f8c62012-04-25 16:51:00 +0100279 return;
Stefan Hajnoczi4f1043b2012-01-18 14:40:44 +0000280 }
281
282 s->base = base;
Marcelo Tosattic8c30802012-01-18 14:40:53 +0000283 if (base_id) {
284 pstrcpy(s->backing_file_id, sizeof(s->backing_file_id), base_id);
285 }
Stefan Hajnoczi4f1043b2012-01-18 14:40:44 +0000286
Paolo Bonzinifa4478d2012-05-08 16:51:46 +0200287 s->common.co = qemu_coroutine_create(stream_run);
288 trace_stream_start(bs, base, s, s->common.co, opaque);
289 qemu_coroutine_enter(s->common.co, s);
Stefan Hajnoczi4f1043b2012-01-18 14:40:44 +0000290}