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" 154f1043b4SStefan Hajnoczi #include "block_int.h" 164f1043b4SStefan Hajnoczi 174f1043b4SStefan Hajnoczi enum { 184f1043b4SStefan Hajnoczi /* 194f1043b4SStefan Hajnoczi * Size of data buffer for populating the image file. This should be large 204f1043b4SStefan Hajnoczi * enough to process multiple clusters in a single call, so that populating 214f1043b4SStefan Hajnoczi * contiguous regions of the image is efficient. 224f1043b4SStefan Hajnoczi */ 234f1043b4SStefan Hajnoczi STREAM_BUFFER_SIZE = 512 * 1024, /* in bytes */ 244f1043b4SStefan Hajnoczi }; 254f1043b4SStefan Hajnoczi 265094a6c0SStefan Hajnoczi #define SLICE_TIME 100000000ULL /* ns */ 275094a6c0SStefan Hajnoczi 285094a6c0SStefan Hajnoczi typedef struct { 295094a6c0SStefan Hajnoczi int64_t next_slice_time; 305094a6c0SStefan Hajnoczi uint64_t slice_quota; 315094a6c0SStefan Hajnoczi uint64_t dispatched; 325094a6c0SStefan Hajnoczi } RateLimit; 335094a6c0SStefan Hajnoczi 345094a6c0SStefan Hajnoczi static int64_t ratelimit_calculate_delay(RateLimit *limit, uint64_t n) 355094a6c0SStefan Hajnoczi { 365094a6c0SStefan Hajnoczi int64_t delay_ns = 0; 375094a6c0SStefan Hajnoczi int64_t now = qemu_get_clock_ns(rt_clock); 385094a6c0SStefan Hajnoczi 395094a6c0SStefan Hajnoczi if (limit->next_slice_time < now) { 405094a6c0SStefan Hajnoczi limit->next_slice_time = now + SLICE_TIME; 415094a6c0SStefan Hajnoczi limit->dispatched = 0; 425094a6c0SStefan Hajnoczi } 435094a6c0SStefan Hajnoczi if (limit->dispatched + n > limit->slice_quota) { 445094a6c0SStefan Hajnoczi delay_ns = limit->next_slice_time - now; 455094a6c0SStefan Hajnoczi } else { 465094a6c0SStefan Hajnoczi limit->dispatched += n; 475094a6c0SStefan Hajnoczi } 485094a6c0SStefan Hajnoczi return delay_ns; 495094a6c0SStefan Hajnoczi } 505094a6c0SStefan Hajnoczi 515094a6c0SStefan Hajnoczi static void ratelimit_set_speed(RateLimit *limit, uint64_t speed) 525094a6c0SStefan Hajnoczi { 535094a6c0SStefan Hajnoczi limit->slice_quota = speed / (1000000000ULL / SLICE_TIME); 545094a6c0SStefan Hajnoczi } 555094a6c0SStefan Hajnoczi 564f1043b4SStefan Hajnoczi typedef struct StreamBlockJob { 574f1043b4SStefan Hajnoczi BlockJob common; 585094a6c0SStefan Hajnoczi RateLimit limit; 594f1043b4SStefan Hajnoczi BlockDriverState *base; 60c8c3080fSMarcelo Tosatti char backing_file_id[1024]; 614f1043b4SStefan Hajnoczi } StreamBlockJob; 624f1043b4SStefan Hajnoczi 634f1043b4SStefan Hajnoczi static int coroutine_fn stream_populate(BlockDriverState *bs, 644f1043b4SStefan Hajnoczi int64_t sector_num, int nb_sectors, 654f1043b4SStefan Hajnoczi void *buf) 664f1043b4SStefan Hajnoczi { 674f1043b4SStefan Hajnoczi struct iovec iov = { 684f1043b4SStefan Hajnoczi .iov_base = buf, 694f1043b4SStefan Hajnoczi .iov_len = nb_sectors * BDRV_SECTOR_SIZE, 704f1043b4SStefan Hajnoczi }; 714f1043b4SStefan Hajnoczi QEMUIOVector qiov; 724f1043b4SStefan Hajnoczi 734f1043b4SStefan Hajnoczi qemu_iovec_init_external(&qiov, &iov, 1); 744f1043b4SStefan Hajnoczi 754f1043b4SStefan Hajnoczi /* Copy-on-read the unallocated clusters */ 764f1043b4SStefan Hajnoczi return bdrv_co_copy_on_readv(bs, sector_num, nb_sectors, &qiov); 774f1043b4SStefan Hajnoczi } 784f1043b4SStefan Hajnoczi 795a67a104SMarcelo Tosatti static void close_unused_images(BlockDriverState *top, BlockDriverState *base, 805a67a104SMarcelo Tosatti const char *base_id) 815a67a104SMarcelo Tosatti { 825a67a104SMarcelo Tosatti BlockDriverState *intermediate; 835a67a104SMarcelo Tosatti intermediate = top->backing_hd; 845a67a104SMarcelo Tosatti 855a67a104SMarcelo Tosatti while (intermediate) { 865a67a104SMarcelo Tosatti BlockDriverState *unused; 875a67a104SMarcelo Tosatti 885a67a104SMarcelo Tosatti /* reached base */ 895a67a104SMarcelo Tosatti if (intermediate == base) { 905a67a104SMarcelo Tosatti break; 915a67a104SMarcelo Tosatti } 925a67a104SMarcelo Tosatti 935a67a104SMarcelo Tosatti unused = intermediate; 945a67a104SMarcelo Tosatti intermediate = intermediate->backing_hd; 955a67a104SMarcelo Tosatti unused->backing_hd = NULL; 965a67a104SMarcelo Tosatti bdrv_delete(unused); 975a67a104SMarcelo Tosatti } 985a67a104SMarcelo Tosatti top->backing_hd = base; 995a67a104SMarcelo Tosatti 1005a67a104SMarcelo Tosatti pstrcpy(top->backing_file, sizeof(top->backing_file), ""); 1015a67a104SMarcelo Tosatti pstrcpy(top->backing_format, sizeof(top->backing_format), ""); 1025a67a104SMarcelo Tosatti if (base_id) { 1035a67a104SMarcelo Tosatti pstrcpy(top->backing_file, sizeof(top->backing_file), base_id); 1045a67a104SMarcelo Tosatti if (base->drv) { 1055a67a104SMarcelo Tosatti pstrcpy(top->backing_format, sizeof(top->backing_format), 1065a67a104SMarcelo Tosatti base->drv->format_name); 1075a67a104SMarcelo Tosatti } 1085a67a104SMarcelo Tosatti } 1095a67a104SMarcelo Tosatti 1105a67a104SMarcelo Tosatti } 1115a67a104SMarcelo Tosatti 112c8c3080fSMarcelo Tosatti /* 113c8c3080fSMarcelo Tosatti * Given an image chain: [BASE] -> [INTER1] -> [INTER2] -> [TOP] 114c8c3080fSMarcelo Tosatti * 115c8c3080fSMarcelo Tosatti * Return true if the given sector is allocated in top. 116c8c3080fSMarcelo Tosatti * Return false if the given sector is allocated in intermediate images. 117c8c3080fSMarcelo Tosatti * Return true otherwise. 118c8c3080fSMarcelo Tosatti * 119c8c3080fSMarcelo Tosatti * 'pnum' is set to the number of sectors (including and immediately following 120c8c3080fSMarcelo Tosatti * the specified sector) that are known to be in the same 121c8c3080fSMarcelo Tosatti * allocated/unallocated state. 122c8c3080fSMarcelo Tosatti * 123c8c3080fSMarcelo Tosatti */ 124c8c3080fSMarcelo Tosatti static int coroutine_fn is_allocated_base(BlockDriverState *top, 125c8c3080fSMarcelo Tosatti BlockDriverState *base, 126c8c3080fSMarcelo Tosatti int64_t sector_num, 127c8c3080fSMarcelo Tosatti int nb_sectors, int *pnum) 128c8c3080fSMarcelo Tosatti { 129c8c3080fSMarcelo Tosatti BlockDriverState *intermediate; 130c8c3080fSMarcelo Tosatti int ret, n; 131c8c3080fSMarcelo Tosatti 132c8c3080fSMarcelo Tosatti ret = bdrv_co_is_allocated(top, sector_num, nb_sectors, &n); 133c8c3080fSMarcelo Tosatti if (ret) { 134c8c3080fSMarcelo Tosatti *pnum = n; 135c8c3080fSMarcelo Tosatti return ret; 136c8c3080fSMarcelo Tosatti } 137c8c3080fSMarcelo Tosatti 138c8c3080fSMarcelo Tosatti /* 139c8c3080fSMarcelo Tosatti * Is the unallocated chunk [sector_num, n] also 140c8c3080fSMarcelo Tosatti * unallocated between base and top? 141c8c3080fSMarcelo Tosatti */ 142c8c3080fSMarcelo Tosatti intermediate = top->backing_hd; 143c8c3080fSMarcelo Tosatti 144c8c3080fSMarcelo Tosatti while (intermediate) { 145c8c3080fSMarcelo Tosatti int pnum_inter; 146c8c3080fSMarcelo Tosatti 147c8c3080fSMarcelo Tosatti /* reached base */ 148c8c3080fSMarcelo Tosatti if (intermediate == base) { 149c8c3080fSMarcelo Tosatti *pnum = n; 150c8c3080fSMarcelo Tosatti return 1; 151c8c3080fSMarcelo Tosatti } 152c8c3080fSMarcelo Tosatti ret = bdrv_co_is_allocated(intermediate, sector_num, nb_sectors, 153c8c3080fSMarcelo Tosatti &pnum_inter); 154c8c3080fSMarcelo Tosatti if (ret < 0) { 155c8c3080fSMarcelo Tosatti return ret; 156c8c3080fSMarcelo Tosatti } else if (ret) { 157c8c3080fSMarcelo Tosatti *pnum = pnum_inter; 158c8c3080fSMarcelo Tosatti return 0; 159c8c3080fSMarcelo Tosatti } 160c8c3080fSMarcelo Tosatti 161c8c3080fSMarcelo Tosatti /* 162c8c3080fSMarcelo Tosatti * [sector_num, nb_sectors] is unallocated on top but intermediate 163c8c3080fSMarcelo Tosatti * might have 164c8c3080fSMarcelo Tosatti * 165c8c3080fSMarcelo Tosatti * [sector_num+x, nr_sectors] allocated. 166c8c3080fSMarcelo Tosatti */ 167c8c3080fSMarcelo Tosatti if (n > pnum_inter) { 168c8c3080fSMarcelo Tosatti n = pnum_inter; 169c8c3080fSMarcelo Tosatti } 170c8c3080fSMarcelo Tosatti 171c8c3080fSMarcelo Tosatti intermediate = intermediate->backing_hd; 172c8c3080fSMarcelo Tosatti } 173c8c3080fSMarcelo Tosatti 174c8c3080fSMarcelo Tosatti return 1; 175c8c3080fSMarcelo Tosatti } 176c8c3080fSMarcelo Tosatti 1774f1043b4SStefan Hajnoczi static void coroutine_fn stream_run(void *opaque) 1784f1043b4SStefan Hajnoczi { 1794f1043b4SStefan Hajnoczi StreamBlockJob *s = opaque; 1804f1043b4SStefan Hajnoczi BlockDriverState *bs = s->common.bs; 181c8c3080fSMarcelo Tosatti BlockDriverState *base = s->base; 1824f1043b4SStefan Hajnoczi int64_t sector_num, end; 1834f1043b4SStefan Hajnoczi int ret = 0; 1844f1043b4SStefan Hajnoczi int n; 1854f1043b4SStefan Hajnoczi void *buf; 1864f1043b4SStefan Hajnoczi 1874f1043b4SStefan Hajnoczi s->common.len = bdrv_getlength(bs); 1884f1043b4SStefan Hajnoczi if (s->common.len < 0) { 1894f1043b4SStefan Hajnoczi block_job_complete(&s->common, s->common.len); 1904f1043b4SStefan Hajnoczi return; 1914f1043b4SStefan Hajnoczi } 1924f1043b4SStefan Hajnoczi 1934f1043b4SStefan Hajnoczi end = s->common.len >> BDRV_SECTOR_BITS; 1944f1043b4SStefan Hajnoczi buf = qemu_blockalign(bs, STREAM_BUFFER_SIZE); 1954f1043b4SStefan Hajnoczi 1964f1043b4SStefan Hajnoczi /* Turn on copy-on-read for the whole block device so that guest read 1974f1043b4SStefan Hajnoczi * requests help us make progress. Only do this when copying the entire 1984f1043b4SStefan Hajnoczi * backing chain since the copy-on-read operation does not take base into 1994f1043b4SStefan Hajnoczi * account. 2004f1043b4SStefan Hajnoczi */ 2014f1043b4SStefan Hajnoczi if (!base) { 2024f1043b4SStefan Hajnoczi bdrv_enable_copy_on_read(bs); 2034f1043b4SStefan Hajnoczi } 2044f1043b4SStefan Hajnoczi 2054f1043b4SStefan Hajnoczi for (sector_num = 0; sector_num < end; sector_num += n) { 2065094a6c0SStefan Hajnoczi retry: 2074f1043b4SStefan Hajnoczi if (block_job_is_cancelled(&s->common)) { 2084f1043b4SStefan Hajnoczi break; 2094f1043b4SStefan Hajnoczi } 2104f1043b4SStefan Hajnoczi 2113e914655SPaolo Bonzini s->common.busy = true; 212c8c3080fSMarcelo Tosatti if (base) { 213c8c3080fSMarcelo Tosatti ret = is_allocated_base(bs, base, sector_num, 2144f1043b4SStefan Hajnoczi STREAM_BUFFER_SIZE / BDRV_SECTOR_SIZE, &n); 215c8c3080fSMarcelo Tosatti } else { 216c8c3080fSMarcelo Tosatti ret = bdrv_co_is_allocated(bs, sector_num, 217c8c3080fSMarcelo Tosatti STREAM_BUFFER_SIZE / BDRV_SECTOR_SIZE, 218c8c3080fSMarcelo Tosatti &n); 219c8c3080fSMarcelo Tosatti } 2204f1043b4SStefan Hajnoczi trace_stream_one_iteration(s, sector_num, n, ret); 2214f1043b4SStefan Hajnoczi if (ret == 0) { 2225094a6c0SStefan Hajnoczi if (s->common.speed) { 2235094a6c0SStefan Hajnoczi uint64_t delay_ns = ratelimit_calculate_delay(&s->limit, n); 2245094a6c0SStefan Hajnoczi if (delay_ns > 0) { 2253e914655SPaolo Bonzini s->common.busy = false; 2265094a6c0SStefan Hajnoczi co_sleep_ns(rt_clock, delay_ns); 2275094a6c0SStefan Hajnoczi 2285094a6c0SStefan Hajnoczi /* Recheck cancellation and that sectors are unallocated */ 2295094a6c0SStefan Hajnoczi goto retry; 2305094a6c0SStefan Hajnoczi } 2315094a6c0SStefan Hajnoczi } 2324f1043b4SStefan Hajnoczi ret = stream_populate(bs, sector_num, n, buf); 2334f1043b4SStefan Hajnoczi } 2344f1043b4SStefan Hajnoczi if (ret < 0) { 2354f1043b4SStefan Hajnoczi break; 2364f1043b4SStefan Hajnoczi } 237c8c3080fSMarcelo Tosatti ret = 0; 2384f1043b4SStefan Hajnoczi 2394f1043b4SStefan Hajnoczi /* Publish progress */ 2404f1043b4SStefan Hajnoczi s->common.offset += n * BDRV_SECTOR_SIZE; 2415094a6c0SStefan Hajnoczi 2425094a6c0SStefan Hajnoczi /* Note that even when no rate limit is applied we need to yield 2435094a6c0SStefan Hajnoczi * with no pending I/O here so that qemu_aio_flush() returns. 2445094a6c0SStefan Hajnoczi */ 2453e914655SPaolo Bonzini s->common.busy = false; 2465094a6c0SStefan Hajnoczi co_sleep_ns(rt_clock, 0); 2474f1043b4SStefan Hajnoczi } 2484f1043b4SStefan Hajnoczi 2494f1043b4SStefan Hajnoczi if (!base) { 2504f1043b4SStefan Hajnoczi bdrv_disable_copy_on_read(bs); 2514f1043b4SStefan Hajnoczi } 2524f1043b4SStefan Hajnoczi 2533e914655SPaolo Bonzini if (!block_job_is_cancelled(&s->common) && sector_num == end && ret == 0) { 254c8c3080fSMarcelo Tosatti const char *base_id = NULL; 255c8c3080fSMarcelo Tosatti if (base) { 256c8c3080fSMarcelo Tosatti base_id = s->backing_file_id; 257c8c3080fSMarcelo Tosatti } 258c8c3080fSMarcelo Tosatti ret = bdrv_change_backing_file(bs, base_id, NULL); 2595a67a104SMarcelo Tosatti close_unused_images(bs, base, base_id); 2604f1043b4SStefan Hajnoczi } 2614f1043b4SStefan Hajnoczi 2624f1043b4SStefan Hajnoczi qemu_vfree(buf); 2634f1043b4SStefan Hajnoczi block_job_complete(&s->common, ret); 2644f1043b4SStefan Hajnoczi } 2654f1043b4SStefan Hajnoczi 266*9e6636c7SStefan Hajnoczi static void stream_set_speed(BlockJob *job, int64_t value, Error **errp) 2675094a6c0SStefan Hajnoczi { 2685094a6c0SStefan Hajnoczi StreamBlockJob *s = container_of(job, StreamBlockJob, common); 2695094a6c0SStefan Hajnoczi 2705094a6c0SStefan Hajnoczi if (value < 0) { 271*9e6636c7SStefan Hajnoczi error_set(errp, QERR_INVALID_PARAMETER, "value"); 272*9e6636c7SStefan Hajnoczi return; 2735094a6c0SStefan Hajnoczi } 2745094a6c0SStefan Hajnoczi ratelimit_set_speed(&s->limit, value / BDRV_SECTOR_SIZE); 2755094a6c0SStefan Hajnoczi } 2765094a6c0SStefan Hajnoczi 2774f1043b4SStefan Hajnoczi static BlockJobType stream_job_type = { 2784f1043b4SStefan Hajnoczi .instance_size = sizeof(StreamBlockJob), 2794f1043b4SStefan Hajnoczi .job_type = "stream", 2805094a6c0SStefan Hajnoczi .set_speed = stream_set_speed, 2814f1043b4SStefan Hajnoczi }; 2824f1043b4SStefan Hajnoczi 283fd7f8c65SStefan Hajnoczi void stream_start(BlockDriverState *bs, BlockDriverState *base, 284c8c3080fSMarcelo Tosatti const char *base_id, BlockDriverCompletionFunc *cb, 285fd7f8c65SStefan Hajnoczi void *opaque, Error **errp) 2864f1043b4SStefan Hajnoczi { 2874f1043b4SStefan Hajnoczi StreamBlockJob *s; 2884f1043b4SStefan Hajnoczi Coroutine *co; 2894f1043b4SStefan Hajnoczi 290fd7f8c65SStefan Hajnoczi s = block_job_create(&stream_job_type, bs, cb, opaque, errp); 2914f1043b4SStefan Hajnoczi if (!s) { 292fd7f8c65SStefan Hajnoczi return; 2934f1043b4SStefan Hajnoczi } 2944f1043b4SStefan Hajnoczi 2954f1043b4SStefan Hajnoczi s->base = base; 296c8c3080fSMarcelo Tosatti if (base_id) { 297c8c3080fSMarcelo Tosatti pstrcpy(s->backing_file_id, sizeof(s->backing_file_id), base_id); 298c8c3080fSMarcelo Tosatti } 2994f1043b4SStefan Hajnoczi 3004f1043b4SStefan Hajnoczi co = qemu_coroutine_create(stream_run); 3014f1043b4SStefan Hajnoczi trace_stream_start(bs, base, s, co, opaque); 3024f1043b4SStefan Hajnoczi qemu_coroutine_enter(co, s); 3034f1043b4SStefan Hajnoczi } 304