1 /* 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 "qemu/osdep.h" 15 #include "trace.h" 16 #include "block/block_int.h" 17 #include "block/blockjob_int.h" 18 #include "qapi/error.h" 19 #include "qapi/qmp/qerror.h" 20 #include "qapi/qmp/qdict.h" 21 #include "qemu/ratelimit.h" 22 #include "sysemu/block-backend.h" 23 #include "block/copy-on-read.h" 24 25 enum { 26 /* 27 * Maximum chunk size to feed to copy-on-read. This should be 28 * large enough to process multiple clusters in a single call, so 29 * that populating contiguous regions of the image is efficient. 30 */ 31 STREAM_CHUNK = 512 * 1024, /* in bytes */ 32 }; 33 34 typedef struct StreamBlockJob { 35 BlockJob common; 36 BlockBackend *blk; 37 BlockDriverState *base_overlay; /* COW overlay (stream from this) */ 38 BlockDriverState *above_base; /* Node directly above the base */ 39 BlockDriverState *cor_filter_bs; 40 BlockDriverState *target_bs; 41 BlockdevOnError on_error; 42 char *backing_file_str; 43 bool bs_read_only; 44 } StreamBlockJob; 45 46 static int coroutine_fn stream_populate(BlockBackend *blk, 47 int64_t offset, uint64_t bytes) 48 { 49 assert(bytes < SIZE_MAX); 50 51 return blk_co_preadv(blk, offset, bytes, NULL, BDRV_REQ_PREFETCH); 52 } 53 54 static int stream_prepare(Job *job) 55 { 56 StreamBlockJob *s = container_of(job, StreamBlockJob, common.job); 57 BlockDriverState *unfiltered_bs = bdrv_skip_filters(s->target_bs); 58 BlockDriverState *base; 59 BlockDriverState *unfiltered_base; 60 Error *local_err = NULL; 61 int ret = 0; 62 63 /* We should drop filter at this point, as filter hold the backing chain */ 64 bdrv_cor_filter_drop(s->cor_filter_bs); 65 s->cor_filter_bs = NULL; 66 67 /* 68 * bdrv_set_backing_hd() requires that unfiltered_bs is drained. Drain 69 * already here and use bdrv_set_backing_hd_drained() instead because 70 * the polling during drained_begin() might change the graph, and if we do 71 * this only later, we may end up working with the wrong base node (or it 72 * might even have gone away by the time we want to use it). 73 */ 74 bdrv_drained_begin(unfiltered_bs); 75 76 base = bdrv_filter_or_cow_bs(s->above_base); 77 unfiltered_base = bdrv_skip_filters(base); 78 79 if (bdrv_cow_child(unfiltered_bs)) { 80 const char *base_id = NULL, *base_fmt = NULL; 81 if (unfiltered_base) { 82 base_id = s->backing_file_str ?: unfiltered_base->filename; 83 if (unfiltered_base->drv) { 84 base_fmt = unfiltered_base->drv->format_name; 85 } 86 } 87 88 bdrv_set_backing_hd_drained(unfiltered_bs, base, &local_err); 89 90 /* 91 * This call will do I/O, so the graph can change again from here on. 92 * We have already completed the graph change, so we are not in danger 93 * of operating on the wrong node any more if this happens. 94 */ 95 ret = bdrv_change_backing_file(unfiltered_bs, base_id, base_fmt, false); 96 if (local_err) { 97 error_report_err(local_err); 98 ret = -EPERM; 99 goto out; 100 } 101 } 102 103 out: 104 bdrv_drained_end(unfiltered_bs); 105 return ret; 106 } 107 108 static void stream_clean(Job *job) 109 { 110 StreamBlockJob *s = container_of(job, StreamBlockJob, common.job); 111 112 if (s->cor_filter_bs) { 113 bdrv_cor_filter_drop(s->cor_filter_bs); 114 s->cor_filter_bs = NULL; 115 } 116 117 blk_unref(s->blk); 118 s->blk = NULL; 119 120 /* Reopen the image back in read-only mode if necessary */ 121 if (s->bs_read_only) { 122 /* Give up write permissions before making it read-only */ 123 bdrv_reopen_set_read_only(s->target_bs, true, NULL); 124 } 125 126 g_free(s->backing_file_str); 127 } 128 129 static int coroutine_fn stream_run(Job *job, Error **errp) 130 { 131 StreamBlockJob *s = container_of(job, StreamBlockJob, common.job); 132 BlockDriverState *unfiltered_bs = bdrv_skip_filters(s->target_bs); 133 int64_t len; 134 int64_t offset = 0; 135 uint64_t delay_ns = 0; 136 int error = 0; 137 int64_t n = 0; /* bytes */ 138 139 if (unfiltered_bs == s->base_overlay) { 140 /* Nothing to stream */ 141 return 0; 142 } 143 144 WITH_GRAPH_RDLOCK_GUARD() { 145 len = bdrv_co_getlength(s->target_bs); 146 if (len < 0) { 147 return len; 148 } 149 } 150 job_progress_set_remaining(&s->common.job, len); 151 152 for ( ; offset < len; offset += n) { 153 bool copy; 154 int ret; 155 156 /* Note that even when no rate limit is applied we need to yield 157 * with no pending I/O here so that bdrv_drain_all() returns. 158 */ 159 job_sleep_ns(&s->common.job, delay_ns); 160 if (job_is_cancelled(&s->common.job)) { 161 break; 162 } 163 164 copy = false; 165 166 WITH_GRAPH_RDLOCK_GUARD() { 167 ret = bdrv_is_allocated(unfiltered_bs, offset, STREAM_CHUNK, &n); 168 if (ret == 1) { 169 /* Allocated in the top, no need to copy. */ 170 } else if (ret >= 0) { 171 /* 172 * Copy if allocated in the intermediate images. Limit to the 173 * known-unallocated area [offset, offset+n*BDRV_SECTOR_SIZE). 174 */ 175 ret = bdrv_is_allocated_above(bdrv_cow_bs(unfiltered_bs), 176 s->base_overlay, true, 177 offset, n, &n); 178 /* Finish early if end of backing file has been reached */ 179 if (ret == 0 && n == 0) { 180 n = len - offset; 181 } 182 183 copy = (ret > 0); 184 } 185 } 186 trace_stream_one_iteration(s, offset, n, ret); 187 if (copy) { 188 ret = stream_populate(s->blk, offset, n); 189 } 190 if (ret < 0) { 191 BlockErrorAction action = 192 block_job_error_action(&s->common, s->on_error, true, -ret); 193 if (action == BLOCK_ERROR_ACTION_STOP) { 194 n = 0; 195 continue; 196 } 197 if (error == 0) { 198 error = ret; 199 } 200 if (action == BLOCK_ERROR_ACTION_REPORT) { 201 break; 202 } 203 } 204 205 /* Publish progress */ 206 job_progress_update(&s->common.job, n); 207 if (copy) { 208 delay_ns = block_job_ratelimit_get_delay(&s->common, n); 209 } else { 210 delay_ns = 0; 211 } 212 } 213 214 /* Do not remove the backing file if an error was there but ignored. */ 215 return error; 216 } 217 218 static const BlockJobDriver stream_job_driver = { 219 .job_driver = { 220 .instance_size = sizeof(StreamBlockJob), 221 .job_type = JOB_TYPE_STREAM, 222 .free = block_job_free, 223 .run = stream_run, 224 .prepare = stream_prepare, 225 .clean = stream_clean, 226 .user_resume = block_job_user_resume, 227 }, 228 }; 229 230 void stream_start(const char *job_id, BlockDriverState *bs, 231 BlockDriverState *base, const char *backing_file_str, 232 BlockDriverState *bottom, 233 int creation_flags, int64_t speed, 234 BlockdevOnError on_error, 235 const char *filter_node_name, 236 Error **errp) 237 { 238 StreamBlockJob *s = NULL; 239 BlockDriverState *iter; 240 bool bs_read_only; 241 int basic_flags = BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED; 242 BlockDriverState *base_overlay; 243 BlockDriverState *cor_filter_bs = NULL; 244 BlockDriverState *above_base; 245 QDict *opts; 246 int ret; 247 248 GLOBAL_STATE_CODE(); 249 250 assert(!(base && bottom)); 251 assert(!(backing_file_str && bottom)); 252 253 if (bottom) { 254 /* 255 * New simple interface. The code is written in terms of old interface 256 * with @base parameter (still, it doesn't freeze link to base, so in 257 * this mean old code is correct for new interface). So, for now, just 258 * emulate base_overlay and above_base. Still, when old interface 259 * finally removed, we should refactor code to use only "bottom", but 260 * not "*base*" things. 261 */ 262 assert(!bottom->drv->is_filter); 263 base_overlay = above_base = bottom; 264 } else { 265 base_overlay = bdrv_find_overlay(bs, base); 266 if (!base_overlay) { 267 error_setg(errp, "'%s' is not in the backing chain of '%s'", 268 base->node_name, bs->node_name); 269 return; 270 } 271 272 /* 273 * Find the node directly above @base. @base_overlay is a COW overlay, 274 * so it must have a bdrv_cow_child(), but it is the immediate overlay 275 * of @base, so between the two there can only be filters. 276 */ 277 above_base = base_overlay; 278 if (bdrv_cow_bs(above_base) != base) { 279 above_base = bdrv_cow_bs(above_base); 280 while (bdrv_filter_bs(above_base) != base) { 281 above_base = bdrv_filter_bs(above_base); 282 } 283 } 284 } 285 286 /* Make sure that the image is opened in read-write mode */ 287 bs_read_only = bdrv_is_read_only(bs); 288 if (bs_read_only) { 289 int ret; 290 /* Hold the chain during reopen */ 291 if (bdrv_freeze_backing_chain(bs, above_base, errp) < 0) { 292 return; 293 } 294 295 ret = bdrv_reopen_set_read_only(bs, false, errp); 296 297 /* failure, or cor-filter will hold the chain */ 298 bdrv_unfreeze_backing_chain(bs, above_base); 299 300 if (ret < 0) { 301 return; 302 } 303 } 304 305 opts = qdict_new(); 306 307 qdict_put_str(opts, "driver", "copy-on-read"); 308 qdict_put_str(opts, "file", bdrv_get_node_name(bs)); 309 /* Pass the base_overlay node name as 'bottom' to COR driver */ 310 qdict_put_str(opts, "bottom", base_overlay->node_name); 311 if (filter_node_name) { 312 qdict_put_str(opts, "node-name", filter_node_name); 313 } 314 315 cor_filter_bs = bdrv_insert_node(bs, opts, BDRV_O_RDWR, errp); 316 if (!cor_filter_bs) { 317 goto fail; 318 } 319 320 if (!filter_node_name) { 321 cor_filter_bs->implicit = true; 322 } 323 324 s = block_job_create(job_id, &stream_job_driver, NULL, cor_filter_bs, 325 0, BLK_PERM_ALL, 326 speed, creation_flags, NULL, NULL, errp); 327 if (!s) { 328 goto fail; 329 } 330 331 s->blk = blk_new_with_bs(cor_filter_bs, BLK_PERM_CONSISTENT_READ, 332 basic_flags | BLK_PERM_WRITE, errp); 333 if (!s->blk) { 334 goto fail; 335 } 336 /* 337 * Disable request queuing in the BlockBackend to avoid deadlocks on drain: 338 * The job reports that it's busy until it reaches a pause point. 339 */ 340 blk_set_disable_request_queuing(s->blk, true); 341 blk_set_allow_aio_context_change(s->blk, true); 342 343 /* 344 * Prevent concurrent jobs trying to modify the graph structure here, we 345 * already have our own plans. Also don't allow resize as the image size is 346 * queried only at the job start and then cached. 347 */ 348 if (block_job_add_bdrv(&s->common, "active node", bs, 0, 349 basic_flags | BLK_PERM_WRITE, errp)) { 350 goto fail; 351 } 352 353 /* Block all intermediate nodes between bs and base, because they will 354 * disappear from the chain after this operation. The streaming job reads 355 * every block only once, assuming that it doesn't change, so forbid writes 356 * and resizes. Reassign the base node pointer because the backing BS of the 357 * bottom node might change after the call to bdrv_reopen_set_read_only() 358 * due to parallel block jobs running. 359 * above_base node might change after the call to 360 * bdrv_reopen_set_read_only() due to parallel block jobs running. 361 */ 362 base = bdrv_filter_or_cow_bs(above_base); 363 for (iter = bdrv_filter_or_cow_bs(bs); iter != base; 364 iter = bdrv_filter_or_cow_bs(iter)) 365 { 366 ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0, 367 basic_flags, errp); 368 if (ret < 0) { 369 goto fail; 370 } 371 } 372 373 s->base_overlay = base_overlay; 374 s->above_base = above_base; 375 s->backing_file_str = g_strdup(backing_file_str); 376 s->cor_filter_bs = cor_filter_bs; 377 s->target_bs = bs; 378 s->bs_read_only = bs_read_only; 379 380 s->on_error = on_error; 381 trace_stream_start(bs, base, s); 382 job_start(&s->common.job); 383 return; 384 385 fail: 386 if (s) { 387 job_early_fail(&s->common.job); 388 } 389 if (cor_filter_bs) { 390 bdrv_cor_filter_drop(cor_filter_bs); 391 } 392 if (bs_read_only) { 393 bdrv_reopen_set_read_only(bs, true, NULL); 394 } 395 } 396