14f1043b4SStefan Hajnoczi /* 24f1043b4SStefan Hajnoczi * Image streaming 34f1043b4SStefan Hajnoczi * 44f1043b4SStefan Hajnoczi * Copyright IBM, Corp. 2011 54f1043b4SStefan Hajnoczi * 64f1043b4SStefan Hajnoczi * Authors: 74f1043b4SStefan Hajnoczi * Stefan Hajnoczi <stefanha@linux.vnet.ibm.com> 84f1043b4SStefan Hajnoczi * 94f1043b4SStefan Hajnoczi * This work is licensed under the terms of the GNU LGPL, version 2 or later. 104f1043b4SStefan Hajnoczi * See the COPYING.LIB file in the top-level directory. 114f1043b4SStefan Hajnoczi * 124f1043b4SStefan Hajnoczi */ 134f1043b4SStefan Hajnoczi 144f1043b4SStefan Hajnoczi #include "trace.h" 15737e150eSPaolo Bonzini #include "block/block_int.h" 16737e150eSPaolo Bonzini #include "block/blockjob.h" 17*cc7a8ea7SMarkus Armbruster #include "qapi/qmp/qerror.h" 186ef228fcSPaolo Bonzini #include "qemu/ratelimit.h" 194f1043b4SStefan Hajnoczi 204f1043b4SStefan Hajnoczi enum { 214f1043b4SStefan Hajnoczi /* 224f1043b4SStefan Hajnoczi * Size of data buffer for populating the image file. This should be large 234f1043b4SStefan Hajnoczi * enough to process multiple clusters in a single call, so that populating 244f1043b4SStefan Hajnoczi * contiguous regions of the image is efficient. 254f1043b4SStefan Hajnoczi */ 264f1043b4SStefan Hajnoczi STREAM_BUFFER_SIZE = 512 * 1024, /* in bytes */ 274f1043b4SStefan Hajnoczi }; 284f1043b4SStefan Hajnoczi 295094a6c0SStefan Hajnoczi #define SLICE_TIME 100000000ULL /* ns */ 305094a6c0SStefan Hajnoczi 314f1043b4SStefan Hajnoczi typedef struct StreamBlockJob { 324f1043b4SStefan Hajnoczi BlockJob common; 335094a6c0SStefan Hajnoczi RateLimit limit; 344f1043b4SStefan Hajnoczi BlockDriverState *base; 351d809098SPaolo Bonzini BlockdevOnError on_error; 3613d8cc51SJeff Cody char *backing_file_str; 374f1043b4SStefan Hajnoczi } StreamBlockJob; 384f1043b4SStefan Hajnoczi 394f1043b4SStefan Hajnoczi static int coroutine_fn stream_populate(BlockDriverState *bs, 404f1043b4SStefan Hajnoczi int64_t sector_num, int nb_sectors, 414f1043b4SStefan Hajnoczi void *buf) 424f1043b4SStefan Hajnoczi { 434f1043b4SStefan Hajnoczi struct iovec iov = { 444f1043b4SStefan Hajnoczi .iov_base = buf, 454f1043b4SStefan Hajnoczi .iov_len = nb_sectors * BDRV_SECTOR_SIZE, 464f1043b4SStefan Hajnoczi }; 474f1043b4SStefan Hajnoczi QEMUIOVector qiov; 484f1043b4SStefan Hajnoczi 494f1043b4SStefan Hajnoczi qemu_iovec_init_external(&qiov, &iov, 1); 504f1043b4SStefan Hajnoczi 514f1043b4SStefan Hajnoczi /* Copy-on-read the unallocated clusters */ 524f1043b4SStefan Hajnoczi return bdrv_co_copy_on_readv(bs, sector_num, nb_sectors, &qiov); 534f1043b4SStefan Hajnoczi } 544f1043b4SStefan Hajnoczi 555a67a104SMarcelo Tosatti static void close_unused_images(BlockDriverState *top, BlockDriverState *base, 565a67a104SMarcelo Tosatti const char *base_id) 575a67a104SMarcelo Tosatti { 585a67a104SMarcelo Tosatti BlockDriverState *intermediate; 595a67a104SMarcelo Tosatti intermediate = top->backing_hd; 605a67a104SMarcelo Tosatti 6188266f5aSStefan Hajnoczi /* Must assign before bdrv_delete() to prevent traversing dangling pointer 6288266f5aSStefan Hajnoczi * while we delete backing image instances. 6388266f5aSStefan Hajnoczi */ 64920beae1SFam Zheng bdrv_set_backing_hd(top, base); 6588266f5aSStefan Hajnoczi 665a67a104SMarcelo Tosatti while (intermediate) { 675a67a104SMarcelo Tosatti BlockDriverState *unused; 685a67a104SMarcelo Tosatti 695a67a104SMarcelo Tosatti /* reached base */ 705a67a104SMarcelo Tosatti if (intermediate == base) { 715a67a104SMarcelo Tosatti break; 725a67a104SMarcelo Tosatti } 735a67a104SMarcelo Tosatti 745a67a104SMarcelo Tosatti unused = intermediate; 755a67a104SMarcelo Tosatti intermediate = intermediate->backing_hd; 76920beae1SFam Zheng bdrv_set_backing_hd(unused, NULL); 774f6fd349SFam Zheng bdrv_unref(unused); 785a67a104SMarcelo Tosatti } 79355ef4acSKevin Wolf 803baca891SKevin Wolf bdrv_refresh_limits(top, NULL); 815a67a104SMarcelo Tosatti } 825a67a104SMarcelo Tosatti 83f3e69bebSStefan Hajnoczi typedef struct { 84f3e69bebSStefan Hajnoczi int ret; 85f3e69bebSStefan Hajnoczi bool reached_end; 86f3e69bebSStefan Hajnoczi } StreamCompleteData; 87f3e69bebSStefan Hajnoczi 88f3e69bebSStefan Hajnoczi static void stream_complete(BlockJob *job, void *opaque) 89f3e69bebSStefan Hajnoczi { 90f3e69bebSStefan Hajnoczi StreamBlockJob *s = container_of(job, StreamBlockJob, common); 91f3e69bebSStefan Hajnoczi StreamCompleteData *data = opaque; 92f3e69bebSStefan Hajnoczi BlockDriverState *base = s->base; 93f3e69bebSStefan Hajnoczi 94f3e69bebSStefan Hajnoczi if (!block_job_is_cancelled(&s->common) && data->reached_end && 95f3e69bebSStefan Hajnoczi data->ret == 0) { 96f3e69bebSStefan Hajnoczi const char *base_id = NULL, *base_fmt = NULL; 97f3e69bebSStefan Hajnoczi if (base) { 98f3e69bebSStefan Hajnoczi base_id = s->backing_file_str; 99f3e69bebSStefan Hajnoczi if (base->drv) { 100f3e69bebSStefan Hajnoczi base_fmt = base->drv->format_name; 101f3e69bebSStefan Hajnoczi } 102f3e69bebSStefan Hajnoczi } 103f3e69bebSStefan Hajnoczi data->ret = bdrv_change_backing_file(job->bs, base_id, base_fmt); 104f3e69bebSStefan Hajnoczi close_unused_images(job->bs, base, base_id); 105f3e69bebSStefan Hajnoczi } 106f3e69bebSStefan Hajnoczi 107f3e69bebSStefan Hajnoczi g_free(s->backing_file_str); 108f3e69bebSStefan Hajnoczi block_job_completed(&s->common, data->ret); 109f3e69bebSStefan Hajnoczi g_free(data); 110f3e69bebSStefan Hajnoczi } 111f3e69bebSStefan Hajnoczi 1124f1043b4SStefan Hajnoczi static void coroutine_fn stream_run(void *opaque) 1134f1043b4SStefan Hajnoczi { 1144f1043b4SStefan Hajnoczi StreamBlockJob *s = opaque; 115f3e69bebSStefan Hajnoczi StreamCompleteData *data; 1164f1043b4SStefan Hajnoczi BlockDriverState *bs = s->common.bs; 117c8c3080fSMarcelo Tosatti BlockDriverState *base = s->base; 1184f1043b4SStefan Hajnoczi int64_t sector_num, end; 1191d809098SPaolo Bonzini int error = 0; 1204f1043b4SStefan Hajnoczi int ret = 0; 12104120e3bSAnthony Liguori int n = 0; 1224f1043b4SStefan Hajnoczi void *buf; 1234f1043b4SStefan Hajnoczi 124f4a193e7SMax Reitz if (!bs->backing_hd) { 125f4a193e7SMax Reitz block_job_completed(&s->common, 0); 126f4a193e7SMax Reitz return; 127f4a193e7SMax Reitz } 128f4a193e7SMax Reitz 1294f1043b4SStefan Hajnoczi s->common.len = bdrv_getlength(bs); 1304f1043b4SStefan Hajnoczi if (s->common.len < 0) { 13165f46322SPaolo Bonzini block_job_completed(&s->common, s->common.len); 1324f1043b4SStefan Hajnoczi return; 1334f1043b4SStefan Hajnoczi } 1344f1043b4SStefan Hajnoczi 1354f1043b4SStefan Hajnoczi end = s->common.len >> BDRV_SECTOR_BITS; 1364f1043b4SStefan Hajnoczi buf = qemu_blockalign(bs, STREAM_BUFFER_SIZE); 1374f1043b4SStefan Hajnoczi 1384f1043b4SStefan Hajnoczi /* Turn on copy-on-read for the whole block device so that guest read 1394f1043b4SStefan Hajnoczi * requests help us make progress. Only do this when copying the entire 1404f1043b4SStefan Hajnoczi * backing chain since the copy-on-read operation does not take base into 1414f1043b4SStefan Hajnoczi * account. 1424f1043b4SStefan Hajnoczi */ 1434f1043b4SStefan Hajnoczi if (!base) { 1444f1043b4SStefan Hajnoczi bdrv_enable_copy_on_read(bs); 1454f1043b4SStefan Hajnoczi } 1464f1043b4SStefan Hajnoczi 1474f1043b4SStefan Hajnoczi for (sector_num = 0; sector_num < end; sector_num += n) { 1484513eafeSPaolo Bonzini uint64_t delay_ns = 0; 149f9749f28SPaolo Bonzini bool copy; 1504513eafeSPaolo Bonzini 1514513eafeSPaolo Bonzini wait: 1524513eafeSPaolo Bonzini /* Note that even when no rate limit is applied we need to yield 153c57b6656SKevin Wolf * with no pending I/O here so that bdrv_drain_all() returns. 1544513eafeSPaolo Bonzini */ 1557483d1e5SAlex Bligh block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns); 1564f1043b4SStefan Hajnoczi if (block_job_is_cancelled(&s->common)) { 1574f1043b4SStefan Hajnoczi break; 1584f1043b4SStefan Hajnoczi } 1594f1043b4SStefan Hajnoczi 160c3e4f43aSStefan Weil copy = false; 161c3e4f43aSStefan Weil 162bdad13b9SPaolo Bonzini ret = bdrv_is_allocated(bs, sector_num, 1634f1043b4SStefan Hajnoczi STREAM_BUFFER_SIZE / BDRV_SECTOR_SIZE, &n); 164f9749f28SPaolo Bonzini if (ret == 1) { 165f9749f28SPaolo Bonzini /* Allocated in the top, no need to copy. */ 166d663640cSPaolo Bonzini } else if (ret >= 0) { 167f9749f28SPaolo Bonzini /* Copy if allocated in the intermediate images. Limit to the 168f9749f28SPaolo Bonzini * known-unallocated area [sector_num, sector_num+n). */ 1694f578637SPaolo Bonzini ret = bdrv_is_allocated_above(bs->backing_hd, base, 170188a7bbfSPaolo Bonzini sector_num, n, &n); 171571cd9dcSStefan Hajnoczi 172571cd9dcSStefan Hajnoczi /* Finish early if end of backing file has been reached */ 173571cd9dcSStefan Hajnoczi if (ret == 0 && n == 0) { 174571cd9dcSStefan Hajnoczi n = end - sector_num; 175571cd9dcSStefan Hajnoczi } 176571cd9dcSStefan Hajnoczi 177f9749f28SPaolo Bonzini copy = (ret == 1); 178f9749f28SPaolo Bonzini } 1794f1043b4SStefan Hajnoczi trace_stream_one_iteration(s, sector_num, n, ret); 180c3e4f43aSStefan Weil if (copy) { 1815094a6c0SStefan Hajnoczi if (s->common.speed) { 1824513eafeSPaolo Bonzini delay_ns = ratelimit_calculate_delay(&s->limit, n); 1835094a6c0SStefan Hajnoczi if (delay_ns > 0) { 1844513eafeSPaolo Bonzini goto wait; 1855094a6c0SStefan Hajnoczi } 1865094a6c0SStefan Hajnoczi } 1874f1043b4SStefan Hajnoczi ret = stream_populate(bs, sector_num, n, buf); 1884f1043b4SStefan Hajnoczi } 1894f1043b4SStefan Hajnoczi if (ret < 0) { 1901d809098SPaolo Bonzini BlockErrorAction action = 1911d809098SPaolo Bonzini block_job_error_action(&s->common, s->common.bs, s->on_error, 1921d809098SPaolo Bonzini true, -ret); 193a589569fSWenchao Xia if (action == BLOCK_ERROR_ACTION_STOP) { 1941d809098SPaolo Bonzini n = 0; 1951d809098SPaolo Bonzini continue; 1961d809098SPaolo Bonzini } 1971d809098SPaolo Bonzini if (error == 0) { 1981d809098SPaolo Bonzini error = ret; 1991d809098SPaolo Bonzini } 200a589569fSWenchao Xia if (action == BLOCK_ERROR_ACTION_REPORT) { 2014f1043b4SStefan Hajnoczi break; 2024f1043b4SStefan Hajnoczi } 2031d809098SPaolo Bonzini } 204c8c3080fSMarcelo Tosatti ret = 0; 2054f1043b4SStefan Hajnoczi 2064f1043b4SStefan Hajnoczi /* Publish progress */ 2074f1043b4SStefan Hajnoczi s->common.offset += n * BDRV_SECTOR_SIZE; 2084f1043b4SStefan Hajnoczi } 2094f1043b4SStefan Hajnoczi 2104f1043b4SStefan Hajnoczi if (!base) { 2114f1043b4SStefan Hajnoczi bdrv_disable_copy_on_read(bs); 2124f1043b4SStefan Hajnoczi } 2134f1043b4SStefan Hajnoczi 2141d809098SPaolo Bonzini /* Do not remove the backing file if an error was there but ignored. */ 2151d809098SPaolo Bonzini ret = error; 2161d809098SPaolo Bonzini 2174f1043b4SStefan Hajnoczi qemu_vfree(buf); 218f3e69bebSStefan Hajnoczi 219f3e69bebSStefan Hajnoczi /* Modify backing chain and close BDSes in main loop */ 220f3e69bebSStefan Hajnoczi data = g_malloc(sizeof(*data)); 221f3e69bebSStefan Hajnoczi data->ret = ret; 222f3e69bebSStefan Hajnoczi data->reached_end = sector_num == end; 223f3e69bebSStefan Hajnoczi block_job_defer_to_main_loop(&s->common, stream_complete, data); 2244f1043b4SStefan Hajnoczi } 2254f1043b4SStefan Hajnoczi 226882ec7ceSStefan Hajnoczi static void stream_set_speed(BlockJob *job, int64_t speed, Error **errp) 2275094a6c0SStefan Hajnoczi { 2285094a6c0SStefan Hajnoczi StreamBlockJob *s = container_of(job, StreamBlockJob, common); 2295094a6c0SStefan Hajnoczi 230882ec7ceSStefan Hajnoczi if (speed < 0) { 231c6bd8c70SMarkus Armbruster error_setg(errp, QERR_INVALID_PARAMETER, "speed"); 2329e6636c7SStefan Hajnoczi return; 2335094a6c0SStefan Hajnoczi } 2346ef228fcSPaolo Bonzini ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME); 2355094a6c0SStefan Hajnoczi } 2365094a6c0SStefan Hajnoczi 2373fc4b10aSFam Zheng static const BlockJobDriver stream_job_driver = { 2384f1043b4SStefan Hajnoczi .instance_size = sizeof(StreamBlockJob), 23979e14bf7SFam Zheng .job_type = BLOCK_JOB_TYPE_STREAM, 2405094a6c0SStefan Hajnoczi .set_speed = stream_set_speed, 2414f1043b4SStefan Hajnoczi }; 2424f1043b4SStefan Hajnoczi 243fd7f8c65SStefan Hajnoczi void stream_start(BlockDriverState *bs, BlockDriverState *base, 24413d8cc51SJeff Cody const char *backing_file_str, int64_t speed, 2451d809098SPaolo Bonzini BlockdevOnError on_error, 246097310b5SMarkus Armbruster BlockCompletionFunc *cb, 247fd7f8c65SStefan Hajnoczi void *opaque, Error **errp) 2484f1043b4SStefan Hajnoczi { 2494f1043b4SStefan Hajnoczi StreamBlockJob *s; 2504f1043b4SStefan Hajnoczi 2511d809098SPaolo Bonzini if ((on_error == BLOCKDEV_ON_ERROR_STOP || 2521d809098SPaolo Bonzini on_error == BLOCKDEV_ON_ERROR_ENOSPC) && 2531d809098SPaolo Bonzini !bdrv_iostatus_is_enabled(bs)) { 254c6bd8c70SMarkus Armbruster error_setg(errp, QERR_INVALID_PARAMETER, "on-error"); 2551d809098SPaolo Bonzini return; 2561d809098SPaolo Bonzini } 2571d809098SPaolo Bonzini 2583fc4b10aSFam Zheng s = block_job_create(&stream_job_driver, bs, speed, cb, opaque, errp); 2594f1043b4SStefan Hajnoczi if (!s) { 260fd7f8c65SStefan Hajnoczi return; 2614f1043b4SStefan Hajnoczi } 2624f1043b4SStefan Hajnoczi 2634f1043b4SStefan Hajnoczi s->base = base; 26413d8cc51SJeff Cody s->backing_file_str = g_strdup(backing_file_str); 2654f1043b4SStefan Hajnoczi 2661d809098SPaolo Bonzini s->on_error = on_error; 267fa4478d5SPaolo Bonzini s->common.co = qemu_coroutine_create(stream_run); 268fa4478d5SPaolo Bonzini trace_stream_start(bs, base, s, s->common.co, opaque); 269fa4478d5SPaolo Bonzini qemu_coroutine_enter(s->common.co, s); 2704f1043b4SStefan Hajnoczi } 271