129ff7890SWen Congyang /* 229ff7890SWen Congyang * Replication Block filter 329ff7890SWen Congyang * 429ff7890SWen Congyang * Copyright (c) 2016 HUAWEI TECHNOLOGIES CO., LTD. 529ff7890SWen Congyang * Copyright (c) 2016 Intel Corporation 629ff7890SWen Congyang * Copyright (c) 2016 FUJITSU LIMITED 729ff7890SWen Congyang * 829ff7890SWen Congyang * Author: 929ff7890SWen Congyang * Wen Congyang <wency@cn.fujitsu.com> 1029ff7890SWen Congyang * 1129ff7890SWen Congyang * This work is licensed under the terms of the GNU GPL, version 2 or later. 1229ff7890SWen Congyang * See the COPYING file in the top-level directory. 1329ff7890SWen Congyang */ 1429ff7890SWen Congyang 1529ff7890SWen Congyang #include "qemu/osdep.h" 160b8fa32fSMarkus Armbruster #include "qemu/module.h" 17922a01a0SMarkus Armbruster #include "qemu/option.h" 1829ff7890SWen Congyang #include "block/nbd.h" 1929ff7890SWen Congyang #include "block/blockjob.h" 2029ff7890SWen Congyang #include "block/block_int.h" 2129ff7890SWen Congyang #include "block/block_backup.h" 2229ff7890SWen Congyang #include "sysemu/block-backend.h" 2329ff7890SWen Congyang #include "qapi/error.h" 243c4e9647SAlberto Garcia #include "qapi/qmp/qdict.h" 25b0262955SPaolo Bonzini #include "block/replication.h" 2629ff7890SWen Congyang 273c76c606SFam Zheng typedef enum { 283c76c606SFam Zheng BLOCK_REPLICATION_NONE, /* block replication is not started */ 293c76c606SFam Zheng BLOCK_REPLICATION_RUNNING, /* block replication is running */ 303c76c606SFam Zheng BLOCK_REPLICATION_FAILOVER, /* failover is running in background */ 313c76c606SFam Zheng BLOCK_REPLICATION_FAILOVER_FAILED, /* failover failed */ 323c76c606SFam Zheng BLOCK_REPLICATION_DONE, /* block replication is done */ 333c76c606SFam Zheng } ReplicationStage; 343c76c606SFam Zheng 3529ff7890SWen Congyang typedef struct BDRVReplicationState { 3629ff7890SWen Congyang ReplicationMode mode; 373c76c606SFam Zheng ReplicationStage stage; 38cc19f177SVladimir Sementsov-Ogievskiy BlockJob *commit_job; 3929ff7890SWen Congyang BdrvChild *hidden_disk; 4029ff7890SWen Congyang BdrvChild *secondary_disk; 41cc19f177SVladimir Sementsov-Ogievskiy BlockJob *backup_job; 4229ff7890SWen Congyang char *top_id; 4329ff7890SWen Congyang ReplicationState *rs; 4429ff7890SWen Congyang Error *blocker; 453c4e9647SAlberto Garcia bool orig_hidden_read_only; 463c4e9647SAlberto Garcia bool orig_secondary_read_only; 4729ff7890SWen Congyang int error; 4829ff7890SWen Congyang } BDRVReplicationState; 4929ff7890SWen Congyang 5029ff7890SWen Congyang static void replication_start(ReplicationState *rs, ReplicationMode mode, 5129ff7890SWen Congyang Error **errp); 5229ff7890SWen Congyang static void replication_do_checkpoint(ReplicationState *rs, Error **errp); 5329ff7890SWen Congyang static void replication_get_error(ReplicationState *rs, Error **errp); 5429ff7890SWen Congyang static void replication_stop(ReplicationState *rs, bool failover, 5529ff7890SWen Congyang Error **errp); 5629ff7890SWen Congyang 5729ff7890SWen Congyang #define REPLICATION_MODE "mode" 5829ff7890SWen Congyang #define REPLICATION_TOP_ID "top-id" 5929ff7890SWen Congyang static QemuOptsList replication_runtime_opts = { 6029ff7890SWen Congyang .name = "replication", 6129ff7890SWen Congyang .head = QTAILQ_HEAD_INITIALIZER(replication_runtime_opts.head), 6229ff7890SWen Congyang .desc = { 6329ff7890SWen Congyang { 6429ff7890SWen Congyang .name = REPLICATION_MODE, 6529ff7890SWen Congyang .type = QEMU_OPT_STRING, 6629ff7890SWen Congyang }, 6729ff7890SWen Congyang { 6829ff7890SWen Congyang .name = REPLICATION_TOP_ID, 6929ff7890SWen Congyang .type = QEMU_OPT_STRING, 7029ff7890SWen Congyang }, 7129ff7890SWen Congyang { /* end of list */ } 7229ff7890SWen Congyang }, 7329ff7890SWen Congyang }; 7429ff7890SWen Congyang 7529ff7890SWen Congyang static ReplicationOps replication_ops = { 7629ff7890SWen Congyang .start = replication_start, 7729ff7890SWen Congyang .checkpoint = replication_do_checkpoint, 7829ff7890SWen Congyang .get_error = replication_get_error, 7929ff7890SWen Congyang .stop = replication_stop, 8029ff7890SWen Congyang }; 8129ff7890SWen Congyang 8229ff7890SWen Congyang static int replication_open(BlockDriverState *bs, QDict *options, 8329ff7890SWen Congyang int flags, Error **errp) 8429ff7890SWen Congyang { 8529ff7890SWen Congyang int ret; 8629ff7890SWen Congyang BDRVReplicationState *s = bs->opaque; 8729ff7890SWen Congyang QemuOpts *opts = NULL; 8829ff7890SWen Congyang const char *mode; 8929ff7890SWen Congyang const char *top_id; 9029ff7890SWen Congyang 9183930780SVladimir Sementsov-Ogievskiy ret = bdrv_open_file_child(NULL, options, "file", bs, errp); 9283930780SVladimir Sementsov-Ogievskiy if (ret < 0) { 9383930780SVladimir Sementsov-Ogievskiy return ret; 944e4bf5c4SKevin Wolf } 954e4bf5c4SKevin Wolf 9629ff7890SWen Congyang ret = -EINVAL; 9729ff7890SWen Congyang opts = qemu_opts_create(&replication_runtime_opts, NULL, 0, &error_abort); 98a5f9b9dfSMarkus Armbruster if (!qemu_opts_absorb_qdict(opts, options, errp)) { 9929ff7890SWen Congyang goto fail; 10029ff7890SWen Congyang } 10129ff7890SWen Congyang 10229ff7890SWen Congyang mode = qemu_opt_get(opts, REPLICATION_MODE); 10329ff7890SWen Congyang if (!mode) { 104dcfe4805SMarkus Armbruster error_setg(errp, "Missing the option mode"); 10529ff7890SWen Congyang goto fail; 10629ff7890SWen Congyang } 10729ff7890SWen Congyang 10829ff7890SWen Congyang if (!strcmp(mode, "primary")) { 10929ff7890SWen Congyang s->mode = REPLICATION_MODE_PRIMARY; 110f4f2539bSChanglong Xie top_id = qemu_opt_get(opts, REPLICATION_TOP_ID); 111f4f2539bSChanglong Xie if (top_id) { 112dcfe4805SMarkus Armbruster error_setg(errp, 113dcfe4805SMarkus Armbruster "The primary side does not support option top-id"); 114f4f2539bSChanglong Xie goto fail; 115f4f2539bSChanglong Xie } 11629ff7890SWen Congyang } else if (!strcmp(mode, "secondary")) { 11729ff7890SWen Congyang s->mode = REPLICATION_MODE_SECONDARY; 11829ff7890SWen Congyang top_id = qemu_opt_get(opts, REPLICATION_TOP_ID); 11929ff7890SWen Congyang s->top_id = g_strdup(top_id); 12029ff7890SWen Congyang if (!s->top_id) { 121dcfe4805SMarkus Armbruster error_setg(errp, "Missing the option top-id"); 12229ff7890SWen Congyang goto fail; 12329ff7890SWen Congyang } 12429ff7890SWen Congyang } else { 125dcfe4805SMarkus Armbruster error_setg(errp, 12629ff7890SWen Congyang "The option mode's value should be primary or secondary"); 12729ff7890SWen Congyang goto fail; 12829ff7890SWen Congyang } 12929ff7890SWen Congyang 13029ff7890SWen Congyang s->rs = replication_new(bs, &replication_ops); 13129ff7890SWen Congyang 13229ff7890SWen Congyang ret = 0; 13329ff7890SWen Congyang 13429ff7890SWen Congyang fail: 13529ff7890SWen Congyang qemu_opts_del(opts); 13629ff7890SWen Congyang return ret; 13729ff7890SWen Congyang } 13829ff7890SWen Congyang 13929ff7890SWen Congyang static void replication_close(BlockDriverState *bs) 14029ff7890SWen Congyang { 14129ff7890SWen Congyang BDRVReplicationState *s = bs->opaque; 14208558e33SStefan Reiter Job *commit_job; 1433ed4f708SEmanuele Giuseppe Esposito GLOBAL_STATE_CODE(); 14429ff7890SWen Congyang 1453c76c606SFam Zheng if (s->stage == BLOCK_REPLICATION_RUNNING) { 14629ff7890SWen Congyang replication_stop(s->rs, false, NULL); 14729ff7890SWen Congyang } 1483c76c606SFam Zheng if (s->stage == BLOCK_REPLICATION_FAILOVER) { 14908558e33SStefan Reiter commit_job = &s->commit_job->job; 15008558e33SStefan Reiter assert(commit_job->aio_context == qemu_get_current_aio_context()); 1514cfb3f05SHanna Reitz job_cancel_sync(commit_job, false); 15250ab0e09SPaolo Bonzini } 15329ff7890SWen Congyang 15429ff7890SWen Congyang if (s->mode == REPLICATION_MODE_SECONDARY) { 15529ff7890SWen Congyang g_free(s->top_id); 15629ff7890SWen Congyang } 15729ff7890SWen Congyang 15829ff7890SWen Congyang replication_remove(s->rs); 15929ff7890SWen Congyang } 16029ff7890SWen Congyang 16137a9051cSChanglong Xie static void replication_child_perm(BlockDriverState *bs, BdrvChild *c, 162bf8e925eSMax Reitz BdrvChildRole role, 163e0995dc3SKevin Wolf BlockReopenQueue *reopen_queue, 16437a9051cSChanglong Xie uint64_t perm, uint64_t shared, 16537a9051cSChanglong Xie uint64_t *nperm, uint64_t *nshared) 16637a9051cSChanglong Xie { 1673b78420bSLukas Straub if (role & BDRV_CHILD_PRIMARY) { 168611e0653SWang Guang *nperm = BLK_PERM_CONSISTENT_READ; 1693b78420bSLukas Straub } else { 1703b78420bSLukas Straub *nperm = 0; 1713b78420bSLukas Straub } 1723b78420bSLukas Straub 173611e0653SWang Guang if ((bs->open_flags & (BDRV_O_INACTIVE | BDRV_O_RDWR)) == BDRV_O_RDWR) { 174611e0653SWang Guang *nperm |= BLK_PERM_WRITE; 175611e0653SWang Guang } 17678ee6bd0SPhilippe Mathieu-Daudé *nshared = BLK_PERM_CONSISTENT_READ 17778ee6bd0SPhilippe Mathieu-Daudé | BLK_PERM_WRITE 17837a9051cSChanglong Xie | BLK_PERM_WRITE_UNCHANGED; 17937a9051cSChanglong Xie return; 18037a9051cSChanglong Xie } 18137a9051cSChanglong Xie 1828ab8140aSKevin Wolf static int64_t coroutine_fn GRAPH_RDLOCK 1838ab8140aSKevin Wolf replication_co_getlength(BlockDriverState *bs) 18429ff7890SWen Congyang { 185c86422c5SEmanuele Giuseppe Esposito return bdrv_co_getlength(bs->file->bs); 18629ff7890SWen Congyang } 18729ff7890SWen Congyang 18829ff7890SWen Congyang static int replication_get_io_status(BDRVReplicationState *s) 18929ff7890SWen Congyang { 1903c76c606SFam Zheng switch (s->stage) { 19129ff7890SWen Congyang case BLOCK_REPLICATION_NONE: 19229ff7890SWen Congyang return -EIO; 19329ff7890SWen Congyang case BLOCK_REPLICATION_RUNNING: 19429ff7890SWen Congyang return 0; 19529ff7890SWen Congyang case BLOCK_REPLICATION_FAILOVER: 19629ff7890SWen Congyang return s->mode == REPLICATION_MODE_PRIMARY ? -EIO : 0; 19729ff7890SWen Congyang case BLOCK_REPLICATION_FAILOVER_FAILED: 19829ff7890SWen Congyang return s->mode == REPLICATION_MODE_PRIMARY ? -EIO : 1; 19929ff7890SWen Congyang case BLOCK_REPLICATION_DONE: 20029ff7890SWen Congyang /* 20129ff7890SWen Congyang * active commit job completes, and active disk and secondary_disk 20229ff7890SWen Congyang * is swapped, so we can operate bs->file directly 20329ff7890SWen Congyang */ 20429ff7890SWen Congyang return s->mode == REPLICATION_MODE_PRIMARY ? -EIO : 0; 20529ff7890SWen Congyang default: 20629ff7890SWen Congyang abort(); 20729ff7890SWen Congyang } 20829ff7890SWen Congyang } 20929ff7890SWen Congyang 21029ff7890SWen Congyang static int replication_return_value(BDRVReplicationState *s, int ret) 21129ff7890SWen Congyang { 21229ff7890SWen Congyang if (s->mode == REPLICATION_MODE_SECONDARY) { 21329ff7890SWen Congyang return ret; 21429ff7890SWen Congyang } 21529ff7890SWen Congyang 21629ff7890SWen Congyang if (ret < 0) { 21729ff7890SWen Congyang s->error = ret; 21829ff7890SWen Congyang ret = 0; 21929ff7890SWen Congyang } 22029ff7890SWen Congyang 22129ff7890SWen Congyang return ret; 22229ff7890SWen Congyang } 22329ff7890SWen Congyang 224b9b10c35SKevin Wolf static int coroutine_fn GRAPH_RDLOCK 225b9b10c35SKevin Wolf replication_co_readv(BlockDriverState *bs, int64_t sector_num, 226b9b10c35SKevin Wolf int remaining_sectors, QEMUIOVector *qiov) 22729ff7890SWen Congyang { 22829ff7890SWen Congyang BDRVReplicationState *s = bs->opaque; 22929ff7890SWen Congyang int ret; 23029ff7890SWen Congyang 23129ff7890SWen Congyang if (s->mode == REPLICATION_MODE_PRIMARY) { 23229ff7890SWen Congyang /* We only use it to forward primary write requests */ 23329ff7890SWen Congyang return -EIO; 23429ff7890SWen Congyang } 23529ff7890SWen Congyang 23629ff7890SWen Congyang ret = replication_get_io_status(s); 23729ff7890SWen Congyang if (ret < 0) { 23829ff7890SWen Congyang return ret; 23929ff7890SWen Congyang } 24029ff7890SWen Congyang 24104a11d87SEric Blake ret = bdrv_co_preadv(bs->file, sector_num * BDRV_SECTOR_SIZE, 24204a11d87SEric Blake remaining_sectors * BDRV_SECTOR_SIZE, qiov, 0); 243e4f9752cSVladimir Sementsov-Ogievskiy 24429ff7890SWen Congyang return replication_return_value(s, ret); 24529ff7890SWen Congyang } 24629ff7890SWen Congyang 247b9b10c35SKevin Wolf static int coroutine_fn GRAPH_RDLOCK 248b9b10c35SKevin Wolf replication_co_writev(BlockDriverState *bs, int64_t sector_num, 249b9b10c35SKevin Wolf int remaining_sectors, QEMUIOVector *qiov, int flags) 25029ff7890SWen Congyang { 25129ff7890SWen Congyang BDRVReplicationState *s = bs->opaque; 25229ff7890SWen Congyang QEMUIOVector hd_qiov; 25329ff7890SWen Congyang uint64_t bytes_done = 0; 25429ff7890SWen Congyang BdrvChild *top = bs->file; 25529ff7890SWen Congyang BdrvChild *base = s->secondary_disk; 25629ff7890SWen Congyang BdrvChild *target; 25751b0a488SEric Blake int ret; 25851b0a488SEric Blake int64_t n; 25929ff7890SWen Congyang 26029ff7890SWen Congyang ret = replication_get_io_status(s); 26129ff7890SWen Congyang if (ret < 0) { 26229ff7890SWen Congyang goto out; 26329ff7890SWen Congyang } 26429ff7890SWen Congyang 26529ff7890SWen Congyang if (ret == 0) { 26604a11d87SEric Blake ret = bdrv_co_pwritev(top, sector_num * BDRV_SECTOR_SIZE, 26704a11d87SEric Blake remaining_sectors * BDRV_SECTOR_SIZE, qiov, 0); 26829ff7890SWen Congyang return replication_return_value(s, ret); 26929ff7890SWen Congyang } 27029ff7890SWen Congyang 27129ff7890SWen Congyang /* 27229ff7890SWen Congyang * Failover failed, only write to active disk if the sectors 27329ff7890SWen Congyang * have already been allocated in active disk/hidden disk. 27429ff7890SWen Congyang */ 27529ff7890SWen Congyang qemu_iovec_init(&hd_qiov, qiov->niov); 27629ff7890SWen Congyang while (remaining_sectors > 0) { 27751b0a488SEric Blake int64_t count; 27851b0a488SEric Blake 279170d3bd3SAndrey Shinkevich ret = bdrv_is_allocated_above(top->bs, base->bs, false, 28051b0a488SEric Blake sector_num * BDRV_SECTOR_SIZE, 28151b0a488SEric Blake remaining_sectors * BDRV_SECTOR_SIZE, 28251b0a488SEric Blake &count); 28329ff7890SWen Congyang if (ret < 0) { 28429ff7890SWen Congyang goto out1; 28529ff7890SWen Congyang } 28629ff7890SWen Congyang 28751b0a488SEric Blake assert(QEMU_IS_ALIGNED(count, BDRV_SECTOR_SIZE)); 28851b0a488SEric Blake n = count >> BDRV_SECTOR_BITS; 28929ff7890SWen Congyang qemu_iovec_reset(&hd_qiov); 29051b0a488SEric Blake qemu_iovec_concat(&hd_qiov, qiov, bytes_done, count); 29129ff7890SWen Congyang 29229ff7890SWen Congyang target = ret ? top : base; 29304a11d87SEric Blake ret = bdrv_co_pwritev(target, sector_num * BDRV_SECTOR_SIZE, 29404a11d87SEric Blake n * BDRV_SECTOR_SIZE, &hd_qiov, 0); 29529ff7890SWen Congyang if (ret < 0) { 29629ff7890SWen Congyang goto out1; 29729ff7890SWen Congyang } 29829ff7890SWen Congyang 29929ff7890SWen Congyang remaining_sectors -= n; 30029ff7890SWen Congyang sector_num += n; 30151b0a488SEric Blake bytes_done += count; 30229ff7890SWen Congyang } 30329ff7890SWen Congyang 30429ff7890SWen Congyang out1: 30529ff7890SWen Congyang qemu_iovec_destroy(&hd_qiov); 30629ff7890SWen Congyang out: 30729ff7890SWen Congyang return ret; 30829ff7890SWen Congyang } 30929ff7890SWen Congyang 3101e12ecfdSLukas Straub static void secondary_do_checkpoint(BlockDriverState *bs, Error **errp) 31129ff7890SWen Congyang { 3121e12ecfdSLukas Straub BDRVReplicationState *s = bs->opaque; 3131e12ecfdSLukas Straub BdrvChild *active_disk = bs->file; 31429ff7890SWen Congyang Error *local_err = NULL; 31529ff7890SWen Congyang int ret; 31629ff7890SWen Congyang 317cc19f177SVladimir Sementsov-Ogievskiy if (!s->backup_job) { 31829ff7890SWen Congyang error_setg(errp, "Backup job was cancelled unexpectedly"); 31929ff7890SWen Congyang return; 32029ff7890SWen Congyang } 32129ff7890SWen Congyang 322cc19f177SVladimir Sementsov-Ogievskiy backup_do_checkpoint(s->backup_job, &local_err); 32329ff7890SWen Congyang if (local_err) { 32429ff7890SWen Congyang error_propagate(errp, local_err); 32529ff7890SWen Congyang return; 32629ff7890SWen Congyang } 32729ff7890SWen Congyang 3281e12ecfdSLukas Straub if (!active_disk->bs->drv) { 329d470ad42SMax Reitz error_setg(errp, "Active disk %s is ejected", 3301e12ecfdSLukas Straub active_disk->bs->node_name); 331d470ad42SMax Reitz return; 332d470ad42SMax Reitz } 333d470ad42SMax Reitz 3341e12ecfdSLukas Straub ret = bdrv_make_empty(active_disk, errp); 33529ff7890SWen Congyang if (ret < 0) { 33629ff7890SWen Congyang return; 33729ff7890SWen Congyang } 33829ff7890SWen Congyang 339d470ad42SMax Reitz if (!s->hidden_disk->bs->drv) { 340d470ad42SMax Reitz error_setg(errp, "Hidden disk %s is ejected", 341d470ad42SMax Reitz s->hidden_disk->bs->node_name); 342d470ad42SMax Reitz return; 343d470ad42SMax Reitz } 344d470ad42SMax Reitz 345c2cf0ecaSLukas Straub ret = bdrv_make_empty(s->hidden_disk, errp); 34629ff7890SWen Congyang if (ret < 0) { 34729ff7890SWen Congyang return; 34829ff7890SWen Congyang } 34929ff7890SWen Congyang } 35029ff7890SWen Congyang 3513c4e9647SAlberto Garcia /* This function is supposed to be called twice: 3523c4e9647SAlberto Garcia * first with writable = true, then with writable = false. 3533c4e9647SAlberto Garcia * The first call puts s->hidden_disk and s->secondary_disk in 3543c4e9647SAlberto Garcia * r/w mode, and the second puts them back in their original state. 3553c4e9647SAlberto Garcia */ 3568dd9006eSPaolo Bonzini static void reopen_backing_file(BlockDriverState *bs, bool writable, 35729ff7890SWen Congyang Error **errp) 35829ff7890SWen Congyang { 3598dd9006eSPaolo Bonzini BDRVReplicationState *s = bs->opaque; 360a990a42bSLukas Straub BdrvChild *hidden_disk, *secondary_disk; 36129ff7890SWen Congyang BlockReopenQueue *reopen_queue = NULL; 36229ff7890SWen Congyang 363a990a42bSLukas Straub /* 364a990a42bSLukas Straub * s->hidden_disk and s->secondary_disk may not be set yet, as they will 365a990a42bSLukas Straub * only be set after the children are writable. 366a990a42bSLukas Straub */ 367a990a42bSLukas Straub hidden_disk = bs->file->bs->backing; 368a990a42bSLukas Straub secondary_disk = hidden_disk->bs->backing; 369a990a42bSLukas Straub 37029ff7890SWen Congyang if (writable) { 371a990a42bSLukas Straub s->orig_hidden_read_only = bdrv_is_read_only(hidden_disk->bs); 372a990a42bSLukas Straub s->orig_secondary_read_only = bdrv_is_read_only(secondary_disk->bs); 37329ff7890SWen Congyang } 37429ff7890SWen Congyang 3753c4e9647SAlberto Garcia if (s->orig_hidden_read_only) { 3763c4e9647SAlberto Garcia QDict *opts = qdict_new(); 3773c4e9647SAlberto Garcia qdict_put_bool(opts, BDRV_OPT_READ_ONLY, !writable); 378a990a42bSLukas Straub reopen_queue = bdrv_reopen_queue(reopen_queue, hidden_disk->bs, 379077e8e20SAlberto Garcia opts, true); 38029ff7890SWen Congyang } 38129ff7890SWen Congyang 3823c4e9647SAlberto Garcia if (s->orig_secondary_read_only) { 3833c4e9647SAlberto Garcia QDict *opts = qdict_new(); 3843c4e9647SAlberto Garcia qdict_put_bool(opts, BDRV_OPT_READ_ONLY, !writable); 385a990a42bSLukas Straub reopen_queue = bdrv_reopen_queue(reopen_queue, secondary_disk->bs, 386077e8e20SAlberto Garcia opts, true); 38729ff7890SWen Congyang } 38829ff7890SWen Congyang 38929ff7890SWen Congyang if (reopen_queue) { 3906cf42ca2SKevin Wolf AioContext *ctx = bdrv_get_aio_context(bs); 3916cf42ca2SKevin Wolf if (ctx != qemu_get_aio_context()) { 3926cf42ca2SKevin Wolf aio_context_release(ctx); 3936cf42ca2SKevin Wolf } 394992861fbSMarkus Armbruster bdrv_reopen_multiple(reopen_queue, errp); 3956cf42ca2SKevin Wolf if (ctx != qemu_get_aio_context()) { 3966cf42ca2SKevin Wolf aio_context_acquire(ctx); 3976cf42ca2SKevin Wolf } 39829ff7890SWen Congyang } 39929ff7890SWen Congyang } 40029ff7890SWen Congyang 4018dd9006eSPaolo Bonzini static void backup_job_cleanup(BlockDriverState *bs) 40229ff7890SWen Congyang { 4038dd9006eSPaolo Bonzini BDRVReplicationState *s = bs->opaque; 40429ff7890SWen Congyang BlockDriverState *top_bs; 40529ff7890SWen Congyang 406e140f4b7SLukas Straub s->backup_job = NULL; 407e140f4b7SLukas Straub 40829ff7890SWen Congyang top_bs = bdrv_lookup_bs(s->top_id, s->top_id, NULL); 40929ff7890SWen Congyang if (!top_bs) { 41029ff7890SWen Congyang return; 41129ff7890SWen Congyang } 41229ff7890SWen Congyang bdrv_op_unblock_all(top_bs, s->blocker); 41329ff7890SWen Congyang error_free(s->blocker); 4148dd9006eSPaolo Bonzini reopen_backing_file(bs, false, NULL); 41529ff7890SWen Congyang } 41629ff7890SWen Congyang 41729ff7890SWen Congyang static void backup_job_completed(void *opaque, int ret) 41829ff7890SWen Congyang { 4198dd9006eSPaolo Bonzini BlockDriverState *bs = opaque; 4208dd9006eSPaolo Bonzini BDRVReplicationState *s = bs->opaque; 42129ff7890SWen Congyang 4223c76c606SFam Zheng if (s->stage != BLOCK_REPLICATION_FAILOVER) { 42329ff7890SWen Congyang /* The backup job is cancelled unexpectedly */ 42429ff7890SWen Congyang s->error = -EIO; 42529ff7890SWen Congyang } 42629ff7890SWen Congyang 4278dd9006eSPaolo Bonzini backup_job_cleanup(bs); 42829ff7890SWen Congyang } 42929ff7890SWen Congyang 43029ff7890SWen Congyang static bool check_top_bs(BlockDriverState *top_bs, BlockDriverState *bs) 43129ff7890SWen Congyang { 43229ff7890SWen Congyang BdrvChild *child; 43329ff7890SWen Congyang 43429ff7890SWen Congyang /* The bs itself is the top_bs */ 43529ff7890SWen Congyang if (top_bs == bs) { 43629ff7890SWen Congyang return true; 43729ff7890SWen Congyang } 43829ff7890SWen Congyang 43929ff7890SWen Congyang /* Iterate over top_bs's children */ 44029ff7890SWen Congyang QLIST_FOREACH(child, &top_bs->children, next) { 44129ff7890SWen Congyang if (child->bs == bs || check_top_bs(child->bs, bs)) { 44229ff7890SWen Congyang return true; 44329ff7890SWen Congyang } 44429ff7890SWen Congyang } 44529ff7890SWen Congyang 44629ff7890SWen Congyang return false; 44729ff7890SWen Congyang } 44829ff7890SWen Congyang 44929ff7890SWen Congyang static void replication_start(ReplicationState *rs, ReplicationMode mode, 45029ff7890SWen Congyang Error **errp) 45129ff7890SWen Congyang { 45229ff7890SWen Congyang BlockDriverState *bs = rs->opaque; 45329ff7890SWen Congyang BDRVReplicationState *s; 45429ff7890SWen Congyang BlockDriverState *top_bs; 455a990a42bSLukas Straub BdrvChild *active_disk, *hidden_disk, *secondary_disk; 45629ff7890SWen Congyang int64_t active_length, hidden_length, disk_length; 45729ff7890SWen Congyang AioContext *aio_context; 45829ff7890SWen Congyang Error *local_err = NULL; 4592c59fd83SVladimir Sementsov-Ogievskiy BackupPerf perf = { .use_copy_range = true, .max_workers = 1 }; 46029ff7890SWen Congyang 46129ff7890SWen Congyang aio_context = bdrv_get_aio_context(bs); 46229ff7890SWen Congyang aio_context_acquire(aio_context); 46329ff7890SWen Congyang s = bs->opaque; 46429ff7890SWen Congyang 46508ddb4ebSLukas Straub if (s->stage == BLOCK_REPLICATION_DONE || 46608ddb4ebSLukas Straub s->stage == BLOCK_REPLICATION_FAILOVER) { 46708ddb4ebSLukas Straub /* 46808ddb4ebSLukas Straub * This case happens when a secondary is promoted to primary. 46908ddb4ebSLukas Straub * Ignore the request because the secondary side of replication 47008ddb4ebSLukas Straub * doesn't have to do anything anymore. 47108ddb4ebSLukas Straub */ 47208ddb4ebSLukas Straub aio_context_release(aio_context); 47308ddb4ebSLukas Straub return; 47408ddb4ebSLukas Straub } 47508ddb4ebSLukas Straub 4763c76c606SFam Zheng if (s->stage != BLOCK_REPLICATION_NONE) { 47729ff7890SWen Congyang error_setg(errp, "Block replication is running or done"); 47829ff7890SWen Congyang aio_context_release(aio_context); 47929ff7890SWen Congyang return; 48029ff7890SWen Congyang } 48129ff7890SWen Congyang 48229ff7890SWen Congyang if (s->mode != mode) { 48329ff7890SWen Congyang error_setg(errp, "The parameter mode's value is invalid, needs %d," 48429ff7890SWen Congyang " but got %d", s->mode, mode); 48529ff7890SWen Congyang aio_context_release(aio_context); 48629ff7890SWen Congyang return; 48729ff7890SWen Congyang } 48829ff7890SWen Congyang 48929ff7890SWen Congyang switch (s->mode) { 49029ff7890SWen Congyang case REPLICATION_MODE_PRIMARY: 49129ff7890SWen Congyang break; 49229ff7890SWen Congyang case REPLICATION_MODE_SECONDARY: 4931e12ecfdSLukas Straub active_disk = bs->file; 4941e12ecfdSLukas Straub if (!active_disk || !active_disk->bs || !active_disk->bs->backing) { 49529ff7890SWen Congyang error_setg(errp, "Active disk doesn't have backing file"); 49629ff7890SWen Congyang aio_context_release(aio_context); 49729ff7890SWen Congyang return; 49829ff7890SWen Congyang } 49929ff7890SWen Congyang 500a990a42bSLukas Straub hidden_disk = active_disk->bs->backing; 501a990a42bSLukas Straub if (!hidden_disk->bs || !hidden_disk->bs->backing) { 50229ff7890SWen Congyang error_setg(errp, "Hidden disk doesn't have backing file"); 50329ff7890SWen Congyang aio_context_release(aio_context); 50429ff7890SWen Congyang return; 50529ff7890SWen Congyang } 50629ff7890SWen Congyang 507a990a42bSLukas Straub secondary_disk = hidden_disk->bs->backing; 508a990a42bSLukas Straub if (!secondary_disk->bs || !bdrv_has_blk(secondary_disk->bs)) { 50929ff7890SWen Congyang error_setg(errp, "The secondary disk doesn't have block backend"); 51029ff7890SWen Congyang aio_context_release(aio_context); 51129ff7890SWen Congyang return; 51229ff7890SWen Congyang } 51329ff7890SWen Congyang 51429ff7890SWen Congyang /* verify the length */ 5151e12ecfdSLukas Straub active_length = bdrv_getlength(active_disk->bs); 516a990a42bSLukas Straub hidden_length = bdrv_getlength(hidden_disk->bs); 517a990a42bSLukas Straub disk_length = bdrv_getlength(secondary_disk->bs); 51829ff7890SWen Congyang if (active_length < 0 || hidden_length < 0 || disk_length < 0 || 51929ff7890SWen Congyang active_length != hidden_length || hidden_length != disk_length) { 52029ff7890SWen Congyang error_setg(errp, "Active disk, hidden disk, secondary disk's length" 52129ff7890SWen Congyang " are not the same"); 52229ff7890SWen Congyang aio_context_release(aio_context); 52329ff7890SWen Congyang return; 52429ff7890SWen Congyang } 52529ff7890SWen Congyang 526d470ad42SMax Reitz /* Must be true, or the bdrv_getlength() calls would have failed */ 527a990a42bSLukas Straub assert(active_disk->bs->drv && hidden_disk->bs->drv); 528d470ad42SMax Reitz 5291e12ecfdSLukas Straub if (!active_disk->bs->drv->bdrv_make_empty || 530a990a42bSLukas Straub !hidden_disk->bs->drv->bdrv_make_empty) { 53129ff7890SWen Congyang error_setg(errp, 53229ff7890SWen Congyang "Active disk or hidden disk doesn't support make_empty"); 53329ff7890SWen Congyang aio_context_release(aio_context); 53429ff7890SWen Congyang return; 53529ff7890SWen Congyang } 53629ff7890SWen Congyang 53729ff7890SWen Congyang /* reopen the backing file in r/w mode */ 5388dd9006eSPaolo Bonzini reopen_backing_file(bs, true, &local_err); 53929ff7890SWen Congyang if (local_err) { 54029ff7890SWen Congyang error_propagate(errp, local_err); 54129ff7890SWen Congyang aio_context_release(aio_context); 54229ff7890SWen Congyang return; 54329ff7890SWen Congyang } 54429ff7890SWen Congyang 545afdaeb9eSKevin Wolf bdrv_graph_wrlock(bs); 546afdaeb9eSKevin Wolf 5473b78420bSLukas Straub bdrv_ref(hidden_disk->bs); 5483b78420bSLukas Straub s->hidden_disk = bdrv_attach_child(bs, hidden_disk->bs, "hidden disk", 5493b78420bSLukas Straub &child_of_bds, BDRV_CHILD_DATA, 5503b78420bSLukas Straub &local_err); 5513b78420bSLukas Straub if (local_err) { 5523b78420bSLukas Straub error_propagate(errp, local_err); 553afdaeb9eSKevin Wolf bdrv_graph_wrunlock(); 5543b78420bSLukas Straub aio_context_release(aio_context); 5553b78420bSLukas Straub return; 5563b78420bSLukas Straub } 5573b78420bSLukas Straub 5583b78420bSLukas Straub bdrv_ref(secondary_disk->bs); 5593b78420bSLukas Straub s->secondary_disk = bdrv_attach_child(bs, secondary_disk->bs, 5603b78420bSLukas Straub "secondary disk", &child_of_bds, 5613b78420bSLukas Straub BDRV_CHILD_DATA, &local_err); 5623b78420bSLukas Straub if (local_err) { 5633b78420bSLukas Straub error_propagate(errp, local_err); 564afdaeb9eSKevin Wolf bdrv_graph_wrunlock(); 5653b78420bSLukas Straub aio_context_release(aio_context); 5663b78420bSLukas Straub return; 5673b78420bSLukas Straub } 568a990a42bSLukas Straub 569afdaeb9eSKevin Wolf bdrv_graph_wrunlock(); 570afdaeb9eSKevin Wolf 57129ff7890SWen Congyang /* start backup job now */ 57229ff7890SWen Congyang error_setg(&s->blocker, 57329ff7890SWen Congyang "Block device is in use by internal backup job"); 57429ff7890SWen Congyang 57529ff7890SWen Congyang top_bs = bdrv_lookup_bs(s->top_id, s->top_id, NULL); 57629ff7890SWen Congyang if (!top_bs || !bdrv_is_root_node(top_bs) || 57729ff7890SWen Congyang !check_top_bs(top_bs, bs)) { 57829ff7890SWen Congyang error_setg(errp, "No top_bs or it is invalid"); 5798dd9006eSPaolo Bonzini reopen_backing_file(bs, false, NULL); 58029ff7890SWen Congyang aio_context_release(aio_context); 58129ff7890SWen Congyang return; 58229ff7890SWen Congyang } 58329ff7890SWen Congyang bdrv_op_block_all(top_bs, s->blocker); 58429ff7890SWen Congyang bdrv_op_unblock(top_bs, BLOCK_OP_TYPE_DATAPLANE, s->blocker); 58529ff7890SWen Congyang 586cc19f177SVladimir Sementsov-Ogievskiy s->backup_job = backup_job_create( 587cc19f177SVladimir Sementsov-Ogievskiy NULL, s->secondary_disk->bs, s->hidden_disk->bs, 58800e30f05SVladimir Sementsov-Ogievskiy 0, MIRROR_SYNC_MODE_NONE, NULL, 0, false, NULL, 58986c6a3b6SVladimir Sementsov-Ogievskiy &perf, 590111049a4SJohn Snow BLOCKDEV_ON_ERROR_REPORT, 591bb02b65cSKevin Wolf BLOCKDEV_ON_ERROR_REPORT, JOB_INTERNAL, 592111049a4SJohn Snow backup_job_completed, bs, NULL, &local_err); 59329ff7890SWen Congyang if (local_err) { 59429ff7890SWen Congyang error_propagate(errp, local_err); 5958dd9006eSPaolo Bonzini backup_job_cleanup(bs); 59629ff7890SWen Congyang aio_context_release(aio_context); 59729ff7890SWen Congyang return; 59829ff7890SWen Congyang } 599cc19f177SVladimir Sementsov-Ogievskiy job_start(&s->backup_job->job); 60029ff7890SWen Congyang break; 60129ff7890SWen Congyang default: 60229ff7890SWen Congyang aio_context_release(aio_context); 60329ff7890SWen Congyang abort(); 60429ff7890SWen Congyang } 60529ff7890SWen Congyang 6063c76c606SFam Zheng s->stage = BLOCK_REPLICATION_RUNNING; 60729ff7890SWen Congyang 60829ff7890SWen Congyang if (s->mode == REPLICATION_MODE_SECONDARY) { 6091e12ecfdSLukas Straub secondary_do_checkpoint(bs, errp); 61029ff7890SWen Congyang } 61129ff7890SWen Congyang 61229ff7890SWen Congyang s->error = 0; 61329ff7890SWen Congyang aio_context_release(aio_context); 61429ff7890SWen Congyang } 61529ff7890SWen Congyang 61629ff7890SWen Congyang static void replication_do_checkpoint(ReplicationState *rs, Error **errp) 61729ff7890SWen Congyang { 61829ff7890SWen Congyang BlockDriverState *bs = rs->opaque; 61929ff7890SWen Congyang BDRVReplicationState *s; 62029ff7890SWen Congyang AioContext *aio_context; 62129ff7890SWen Congyang 62229ff7890SWen Congyang aio_context = bdrv_get_aio_context(bs); 62329ff7890SWen Congyang aio_context_acquire(aio_context); 62429ff7890SWen Congyang s = bs->opaque; 62529ff7890SWen Congyang 62608ddb4ebSLukas Straub if (s->stage == BLOCK_REPLICATION_DONE || 62708ddb4ebSLukas Straub s->stage == BLOCK_REPLICATION_FAILOVER) { 62808ddb4ebSLukas Straub /* 62908ddb4ebSLukas Straub * This case happens when a secondary was promoted to primary. 63008ddb4ebSLukas Straub * Ignore the request because the secondary side of replication 63108ddb4ebSLukas Straub * doesn't have to do anything anymore. 63208ddb4ebSLukas Straub */ 63308ddb4ebSLukas Straub aio_context_release(aio_context); 63408ddb4ebSLukas Straub return; 63508ddb4ebSLukas Straub } 63608ddb4ebSLukas Straub 63729ff7890SWen Congyang if (s->mode == REPLICATION_MODE_SECONDARY) { 6381e12ecfdSLukas Straub secondary_do_checkpoint(bs, errp); 63929ff7890SWen Congyang } 64029ff7890SWen Congyang aio_context_release(aio_context); 64129ff7890SWen Congyang } 64229ff7890SWen Congyang 64329ff7890SWen Congyang static void replication_get_error(ReplicationState *rs, Error **errp) 64429ff7890SWen Congyang { 64529ff7890SWen Congyang BlockDriverState *bs = rs->opaque; 64629ff7890SWen Congyang BDRVReplicationState *s; 64729ff7890SWen Congyang AioContext *aio_context; 64829ff7890SWen Congyang 64929ff7890SWen Congyang aio_context = bdrv_get_aio_context(bs); 65029ff7890SWen Congyang aio_context_acquire(aio_context); 65129ff7890SWen Congyang s = bs->opaque; 65229ff7890SWen Congyang 65308ddb4ebSLukas Straub if (s->stage == BLOCK_REPLICATION_NONE) { 65429ff7890SWen Congyang error_setg(errp, "Block replication is not running"); 65529ff7890SWen Congyang aio_context_release(aio_context); 65629ff7890SWen Congyang return; 65729ff7890SWen Congyang } 65829ff7890SWen Congyang 65929ff7890SWen Congyang if (s->error) { 66029ff7890SWen Congyang error_setg(errp, "I/O error occurred"); 66129ff7890SWen Congyang aio_context_release(aio_context); 66229ff7890SWen Congyang return; 66329ff7890SWen Congyang } 66429ff7890SWen Congyang aio_context_release(aio_context); 66529ff7890SWen Congyang } 66629ff7890SWen Congyang 66729ff7890SWen Congyang static void replication_done(void *opaque, int ret) 66829ff7890SWen Congyang { 66929ff7890SWen Congyang BlockDriverState *bs = opaque; 67029ff7890SWen Congyang BDRVReplicationState *s = bs->opaque; 67129ff7890SWen Congyang 67229ff7890SWen Congyang if (ret == 0) { 6733c76c606SFam Zheng s->stage = BLOCK_REPLICATION_DONE; 67429ff7890SWen Congyang 675*32a8aba3SKevin Wolf bdrv_graph_wrlock(NULL); 6763b78420bSLukas Straub bdrv_unref_child(bs, s->secondary_disk); 67729ff7890SWen Congyang s->secondary_disk = NULL; 6783b78420bSLukas Straub bdrv_unref_child(bs, s->hidden_disk); 67929ff7890SWen Congyang s->hidden_disk = NULL; 680*32a8aba3SKevin Wolf bdrv_graph_wrunlock(); 681*32a8aba3SKevin Wolf 68229ff7890SWen Congyang s->error = 0; 68329ff7890SWen Congyang } else { 6843c76c606SFam Zheng s->stage = BLOCK_REPLICATION_FAILOVER_FAILED; 68529ff7890SWen Congyang s->error = -EIO; 68629ff7890SWen Congyang } 68729ff7890SWen Congyang } 68829ff7890SWen Congyang 68929ff7890SWen Congyang static void replication_stop(ReplicationState *rs, bool failover, Error **errp) 69029ff7890SWen Congyang { 69129ff7890SWen Congyang BlockDriverState *bs = rs->opaque; 69229ff7890SWen Congyang BDRVReplicationState *s; 69329ff7890SWen Congyang AioContext *aio_context; 69429ff7890SWen Congyang 69529ff7890SWen Congyang aio_context = bdrv_get_aio_context(bs); 69629ff7890SWen Congyang aio_context_acquire(aio_context); 69729ff7890SWen Congyang s = bs->opaque; 69829ff7890SWen Congyang 69908ddb4ebSLukas Straub if (s->stage == BLOCK_REPLICATION_DONE || 70008ddb4ebSLukas Straub s->stage == BLOCK_REPLICATION_FAILOVER) { 70108ddb4ebSLukas Straub /* 70208ddb4ebSLukas Straub * This case happens when a secondary was promoted to primary. 70308ddb4ebSLukas Straub * Ignore the request because the secondary side of replication 70408ddb4ebSLukas Straub * doesn't have to do anything anymore. 70508ddb4ebSLukas Straub */ 70608ddb4ebSLukas Straub aio_context_release(aio_context); 70708ddb4ebSLukas Straub return; 70808ddb4ebSLukas Straub } 70908ddb4ebSLukas Straub 7103c76c606SFam Zheng if (s->stage != BLOCK_REPLICATION_RUNNING) { 71129ff7890SWen Congyang error_setg(errp, "Block replication is not running"); 71229ff7890SWen Congyang aio_context_release(aio_context); 71329ff7890SWen Congyang return; 71429ff7890SWen Congyang } 71529ff7890SWen Congyang 71629ff7890SWen Congyang switch (s->mode) { 71729ff7890SWen Congyang case REPLICATION_MODE_PRIMARY: 7183c76c606SFam Zheng s->stage = BLOCK_REPLICATION_DONE; 71929ff7890SWen Congyang s->error = 0; 72029ff7890SWen Congyang break; 72129ff7890SWen Congyang case REPLICATION_MODE_SECONDARY: 72229ff7890SWen Congyang /* 72329ff7890SWen Congyang * This BDS will be closed, and the job should be completed 72429ff7890SWen Congyang * before the BDS is closed, because we will access hidden 72529ff7890SWen Congyang * disk, secondary disk in backup_job_completed(). 72629ff7890SWen Congyang */ 727cc19f177SVladimir Sementsov-Ogievskiy if (s->backup_job) { 7286f592e5aSEmanuele Giuseppe Esposito aio_context_release(aio_context); 7294cfb3f05SHanna Reitz job_cancel_sync(&s->backup_job->job, true); 7306f592e5aSEmanuele Giuseppe Esposito aio_context_acquire(aio_context); 73129ff7890SWen Congyang } 73229ff7890SWen Congyang 73329ff7890SWen Congyang if (!failover) { 7341e12ecfdSLukas Straub secondary_do_checkpoint(bs, errp); 7353c76c606SFam Zheng s->stage = BLOCK_REPLICATION_DONE; 73629ff7890SWen Congyang aio_context_release(aio_context); 73729ff7890SWen Congyang return; 73829ff7890SWen Congyang } 73929ff7890SWen Congyang 7403c76c606SFam Zheng s->stage = BLOCK_REPLICATION_FAILOVER; 741cc19f177SVladimir Sementsov-Ogievskiy s->commit_job = commit_active_start( 7421e12ecfdSLukas Straub NULL, bs->file->bs, s->secondary_disk->bs, 743bb02b65cSKevin Wolf JOB_INTERNAL, 0, BLOCKDEV_ON_ERROR_REPORT, 74478bbd910SFam Zheng NULL, replication_done, bs, true, errp); 74529ff7890SWen Congyang break; 74629ff7890SWen Congyang default: 74729ff7890SWen Congyang aio_context_release(aio_context); 74829ff7890SWen Congyang abort(); 74929ff7890SWen Congyang } 75029ff7890SWen Congyang aio_context_release(aio_context); 75129ff7890SWen Congyang } 75229ff7890SWen Congyang 7532654267cSMax Reitz static const char *const replication_strong_runtime_opts[] = { 7542654267cSMax Reitz REPLICATION_MODE, 7552654267cSMax Reitz REPLICATION_TOP_ID, 7562654267cSMax Reitz 7572654267cSMax Reitz NULL 7582654267cSMax Reitz }; 7592654267cSMax Reitz 760782b9d06SAlberto Garcia static BlockDriver bdrv_replication = { 76129ff7890SWen Congyang .format_name = "replication", 76229ff7890SWen Congyang .instance_size = sizeof(BDRVReplicationState), 76329ff7890SWen Congyang 76429ff7890SWen Congyang .bdrv_open = replication_open, 76529ff7890SWen Congyang .bdrv_close = replication_close, 76637a9051cSChanglong Xie .bdrv_child_perm = replication_child_perm, 76729ff7890SWen Congyang 768c86422c5SEmanuele Giuseppe Esposito .bdrv_co_getlength = replication_co_getlength, 76929ff7890SWen Congyang .bdrv_co_readv = replication_co_readv, 77029ff7890SWen Congyang .bdrv_co_writev = replication_co_writev, 77129ff7890SWen Congyang 77229ff7890SWen Congyang .is_filter = true, 77329ff7890SWen Congyang 7742654267cSMax Reitz .strong_runtime_opts = replication_strong_runtime_opts, 77529ff7890SWen Congyang }; 77629ff7890SWen Congyang 77729ff7890SWen Congyang static void bdrv_replication_init(void) 77829ff7890SWen Congyang { 77929ff7890SWen Congyang bdrv_register(&bdrv_replication); 78029ff7890SWen Congyang } 78129ff7890SWen Congyang 78229ff7890SWen Congyang block_init(bdrv_replication_init); 783