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" 2529ff7890SWen Congyang #include "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; 3829ff7890SWen Congyang BdrvChild *active_disk; 39cc19f177SVladimir Sementsov-Ogievskiy BlockJob *commit_job; 4029ff7890SWen Congyang BdrvChild *hidden_disk; 4129ff7890SWen Congyang BdrvChild *secondary_disk; 42cc19f177SVladimir Sementsov-Ogievskiy BlockJob *backup_job; 4329ff7890SWen Congyang char *top_id; 4429ff7890SWen Congyang ReplicationState *rs; 4529ff7890SWen Congyang Error *blocker; 463c4e9647SAlberto Garcia bool orig_hidden_read_only; 473c4e9647SAlberto Garcia bool orig_secondary_read_only; 4829ff7890SWen Congyang int error; 4929ff7890SWen Congyang } BDRVReplicationState; 5029ff7890SWen Congyang 5129ff7890SWen Congyang static void replication_start(ReplicationState *rs, ReplicationMode mode, 5229ff7890SWen Congyang Error **errp); 5329ff7890SWen Congyang static void replication_do_checkpoint(ReplicationState *rs, Error **errp); 5429ff7890SWen Congyang static void replication_get_error(ReplicationState *rs, Error **errp); 5529ff7890SWen Congyang static void replication_stop(ReplicationState *rs, bool failover, 5629ff7890SWen Congyang Error **errp); 5729ff7890SWen Congyang 5829ff7890SWen Congyang #define REPLICATION_MODE "mode" 5929ff7890SWen Congyang #define REPLICATION_TOP_ID "top-id" 6029ff7890SWen Congyang static QemuOptsList replication_runtime_opts = { 6129ff7890SWen Congyang .name = "replication", 6229ff7890SWen Congyang .head = QTAILQ_HEAD_INITIALIZER(replication_runtime_opts.head), 6329ff7890SWen Congyang .desc = { 6429ff7890SWen Congyang { 6529ff7890SWen Congyang .name = REPLICATION_MODE, 6629ff7890SWen Congyang .type = QEMU_OPT_STRING, 6729ff7890SWen Congyang }, 6829ff7890SWen Congyang { 6929ff7890SWen Congyang .name = REPLICATION_TOP_ID, 7029ff7890SWen Congyang .type = QEMU_OPT_STRING, 7129ff7890SWen Congyang }, 7229ff7890SWen Congyang { /* end of list */ } 7329ff7890SWen Congyang }, 7429ff7890SWen Congyang }; 7529ff7890SWen Congyang 7629ff7890SWen Congyang static ReplicationOps replication_ops = { 7729ff7890SWen Congyang .start = replication_start, 7829ff7890SWen Congyang .checkpoint = replication_do_checkpoint, 7929ff7890SWen Congyang .get_error = replication_get_error, 8029ff7890SWen Congyang .stop = replication_stop, 8129ff7890SWen Congyang }; 8229ff7890SWen Congyang 8329ff7890SWen Congyang static int replication_open(BlockDriverState *bs, QDict *options, 8429ff7890SWen Congyang int flags, Error **errp) 8529ff7890SWen Congyang { 8629ff7890SWen Congyang int ret; 8729ff7890SWen Congyang BDRVReplicationState *s = bs->opaque; 8829ff7890SWen Congyang Error *local_err = NULL; 8929ff7890SWen Congyang QemuOpts *opts = NULL; 9029ff7890SWen Congyang const char *mode; 9129ff7890SWen Congyang const char *top_id; 9229ff7890SWen Congyang 934e4bf5c4SKevin Wolf bs->file = bdrv_open_child(NULL, options, "file", bs, &child_file, 944e4bf5c4SKevin Wolf false, errp); 954e4bf5c4SKevin Wolf if (!bs->file) { 964e4bf5c4SKevin Wolf return -EINVAL; 974e4bf5c4SKevin Wolf } 984e4bf5c4SKevin Wolf 9929ff7890SWen Congyang ret = -EINVAL; 10029ff7890SWen Congyang opts = qemu_opts_create(&replication_runtime_opts, NULL, 0, &error_abort); 10129ff7890SWen Congyang qemu_opts_absorb_qdict(opts, options, &local_err); 10229ff7890SWen Congyang if (local_err) { 10329ff7890SWen Congyang goto fail; 10429ff7890SWen Congyang } 10529ff7890SWen Congyang 10629ff7890SWen Congyang mode = qemu_opt_get(opts, REPLICATION_MODE); 10729ff7890SWen Congyang if (!mode) { 10829ff7890SWen Congyang error_setg(&local_err, "Missing the option mode"); 10929ff7890SWen Congyang goto fail; 11029ff7890SWen Congyang } 11129ff7890SWen Congyang 11229ff7890SWen Congyang if (!strcmp(mode, "primary")) { 11329ff7890SWen Congyang s->mode = REPLICATION_MODE_PRIMARY; 114f4f2539bSChanglong Xie top_id = qemu_opt_get(opts, REPLICATION_TOP_ID); 115f4f2539bSChanglong Xie if (top_id) { 116f4f2539bSChanglong Xie error_setg(&local_err, "The primary side does not support option top-id"); 117f4f2539bSChanglong Xie goto fail; 118f4f2539bSChanglong Xie } 11929ff7890SWen Congyang } else if (!strcmp(mode, "secondary")) { 12029ff7890SWen Congyang s->mode = REPLICATION_MODE_SECONDARY; 12129ff7890SWen Congyang top_id = qemu_opt_get(opts, REPLICATION_TOP_ID); 12229ff7890SWen Congyang s->top_id = g_strdup(top_id); 12329ff7890SWen Congyang if (!s->top_id) { 12429ff7890SWen Congyang error_setg(&local_err, "Missing the option top-id"); 12529ff7890SWen Congyang goto fail; 12629ff7890SWen Congyang } 12729ff7890SWen Congyang } else { 12829ff7890SWen Congyang error_setg(&local_err, 12929ff7890SWen Congyang "The option mode's value should be primary or secondary"); 13029ff7890SWen Congyang goto fail; 13129ff7890SWen Congyang } 13229ff7890SWen Congyang 13329ff7890SWen Congyang s->rs = replication_new(bs, &replication_ops); 13429ff7890SWen Congyang 13529ff7890SWen Congyang ret = 0; 13629ff7890SWen Congyang 13729ff7890SWen Congyang fail: 13829ff7890SWen Congyang qemu_opts_del(opts); 13929ff7890SWen Congyang error_propagate(errp, local_err); 14029ff7890SWen Congyang 14129ff7890SWen Congyang return ret; 14229ff7890SWen Congyang } 14329ff7890SWen Congyang 14429ff7890SWen Congyang static void replication_close(BlockDriverState *bs) 14529ff7890SWen Congyang { 14629ff7890SWen Congyang BDRVReplicationState *s = bs->opaque; 14708558e33SStefan Reiter Job *commit_job; 14829ff7890SWen Congyang 1493c76c606SFam Zheng if (s->stage == BLOCK_REPLICATION_RUNNING) { 15029ff7890SWen Congyang replication_stop(s->rs, false, NULL); 15129ff7890SWen Congyang } 1523c76c606SFam Zheng if (s->stage == BLOCK_REPLICATION_FAILOVER) { 15308558e33SStefan Reiter commit_job = &s->commit_job->job; 15408558e33SStefan Reiter assert(commit_job->aio_context == qemu_get_current_aio_context()); 15508558e33SStefan Reiter job_cancel_sync(commit_job); 15650ab0e09SPaolo Bonzini } 15729ff7890SWen Congyang 15829ff7890SWen Congyang if (s->mode == REPLICATION_MODE_SECONDARY) { 15929ff7890SWen Congyang g_free(s->top_id); 16029ff7890SWen Congyang } 16129ff7890SWen Congyang 16229ff7890SWen Congyang replication_remove(s->rs); 16329ff7890SWen Congyang } 16429ff7890SWen Congyang 16537a9051cSChanglong Xie static void replication_child_perm(BlockDriverState *bs, BdrvChild *c, 166*bd86fb99SMax Reitz const BdrvChildClass *child_class, 167e0995dc3SKevin Wolf BlockReopenQueue *reopen_queue, 16837a9051cSChanglong Xie uint64_t perm, uint64_t shared, 16937a9051cSChanglong Xie uint64_t *nperm, uint64_t *nshared) 17037a9051cSChanglong Xie { 171611e0653SWang Guang *nperm = BLK_PERM_CONSISTENT_READ; 172611e0653SWang Guang if ((bs->open_flags & (BDRV_O_INACTIVE | BDRV_O_RDWR)) == BDRV_O_RDWR) { 173611e0653SWang Guang *nperm |= BLK_PERM_WRITE; 174611e0653SWang Guang } 17578ee6bd0SPhilippe Mathieu-Daudé *nshared = BLK_PERM_CONSISTENT_READ 17678ee6bd0SPhilippe Mathieu-Daudé | BLK_PERM_WRITE 17737a9051cSChanglong Xie | BLK_PERM_WRITE_UNCHANGED; 17837a9051cSChanglong Xie return; 17937a9051cSChanglong Xie } 18037a9051cSChanglong Xie 18129ff7890SWen Congyang static int64_t replication_getlength(BlockDriverState *bs) 18229ff7890SWen Congyang { 18329ff7890SWen Congyang return bdrv_getlength(bs->file->bs); 18429ff7890SWen Congyang } 18529ff7890SWen Congyang 18629ff7890SWen Congyang static int replication_get_io_status(BDRVReplicationState *s) 18729ff7890SWen Congyang { 1883c76c606SFam Zheng switch (s->stage) { 18929ff7890SWen Congyang case BLOCK_REPLICATION_NONE: 19029ff7890SWen Congyang return -EIO; 19129ff7890SWen Congyang case BLOCK_REPLICATION_RUNNING: 19229ff7890SWen Congyang return 0; 19329ff7890SWen Congyang case BLOCK_REPLICATION_FAILOVER: 19429ff7890SWen Congyang return s->mode == REPLICATION_MODE_PRIMARY ? -EIO : 0; 19529ff7890SWen Congyang case BLOCK_REPLICATION_FAILOVER_FAILED: 19629ff7890SWen Congyang return s->mode == REPLICATION_MODE_PRIMARY ? -EIO : 1; 19729ff7890SWen Congyang case BLOCK_REPLICATION_DONE: 19829ff7890SWen Congyang /* 19929ff7890SWen Congyang * active commit job completes, and active disk and secondary_disk 20029ff7890SWen Congyang * is swapped, so we can operate bs->file directly 20129ff7890SWen Congyang */ 20229ff7890SWen Congyang return s->mode == REPLICATION_MODE_PRIMARY ? -EIO : 0; 20329ff7890SWen Congyang default: 20429ff7890SWen Congyang abort(); 20529ff7890SWen Congyang } 20629ff7890SWen Congyang } 20729ff7890SWen Congyang 20829ff7890SWen Congyang static int replication_return_value(BDRVReplicationState *s, int ret) 20929ff7890SWen Congyang { 21029ff7890SWen Congyang if (s->mode == REPLICATION_MODE_SECONDARY) { 21129ff7890SWen Congyang return ret; 21229ff7890SWen Congyang } 21329ff7890SWen Congyang 21429ff7890SWen Congyang if (ret < 0) { 21529ff7890SWen Congyang s->error = ret; 21629ff7890SWen Congyang ret = 0; 21729ff7890SWen Congyang } 21829ff7890SWen Congyang 21929ff7890SWen Congyang return ret; 22029ff7890SWen Congyang } 22129ff7890SWen Congyang 22229ff7890SWen Congyang static coroutine_fn int replication_co_readv(BlockDriverState *bs, 22329ff7890SWen Congyang int64_t sector_num, 22429ff7890SWen Congyang int remaining_sectors, 22529ff7890SWen Congyang QEMUIOVector *qiov) 22629ff7890SWen Congyang { 22729ff7890SWen Congyang BDRVReplicationState *s = bs->opaque; 22829ff7890SWen Congyang int ret; 22929ff7890SWen Congyang 23029ff7890SWen Congyang if (s->mode == REPLICATION_MODE_PRIMARY) { 23129ff7890SWen Congyang /* We only use it to forward primary write requests */ 23229ff7890SWen Congyang return -EIO; 23329ff7890SWen Congyang } 23429ff7890SWen Congyang 23529ff7890SWen Congyang ret = replication_get_io_status(s); 23629ff7890SWen Congyang if (ret < 0) { 23729ff7890SWen Congyang return ret; 23829ff7890SWen Congyang } 23929ff7890SWen Congyang 24004a11d87SEric Blake ret = bdrv_co_preadv(bs->file, sector_num * BDRV_SECTOR_SIZE, 24104a11d87SEric Blake remaining_sectors * BDRV_SECTOR_SIZE, qiov, 0); 242e4f9752cSVladimir Sementsov-Ogievskiy 24329ff7890SWen Congyang return replication_return_value(s, ret); 24429ff7890SWen Congyang } 24529ff7890SWen Congyang 24629ff7890SWen Congyang static coroutine_fn int replication_co_writev(BlockDriverState *bs, 24729ff7890SWen Congyang int64_t sector_num, 24829ff7890SWen Congyang int remaining_sectors, 249e18a58b4SEric Blake QEMUIOVector *qiov, 250e18a58b4SEric Blake int flags) 25129ff7890SWen Congyang { 25229ff7890SWen Congyang BDRVReplicationState *s = bs->opaque; 25329ff7890SWen Congyang QEMUIOVector hd_qiov; 25429ff7890SWen Congyang uint64_t bytes_done = 0; 25529ff7890SWen Congyang BdrvChild *top = bs->file; 25629ff7890SWen Congyang BdrvChild *base = s->secondary_disk; 25729ff7890SWen Congyang BdrvChild *target; 25851b0a488SEric Blake int ret; 25951b0a488SEric Blake int64_t n; 26029ff7890SWen Congyang 261e18a58b4SEric Blake assert(!flags); 26229ff7890SWen Congyang ret = replication_get_io_status(s); 26329ff7890SWen Congyang if (ret < 0) { 26429ff7890SWen Congyang goto out; 26529ff7890SWen Congyang } 26629ff7890SWen Congyang 26729ff7890SWen Congyang if (ret == 0) { 26804a11d87SEric Blake ret = bdrv_co_pwritev(top, sector_num * BDRV_SECTOR_SIZE, 26904a11d87SEric Blake remaining_sectors * BDRV_SECTOR_SIZE, qiov, 0); 27029ff7890SWen Congyang return replication_return_value(s, ret); 27129ff7890SWen Congyang } 27229ff7890SWen Congyang 27329ff7890SWen Congyang /* 27429ff7890SWen Congyang * Failover failed, only write to active disk if the sectors 27529ff7890SWen Congyang * have already been allocated in active disk/hidden disk. 27629ff7890SWen Congyang */ 27729ff7890SWen Congyang qemu_iovec_init(&hd_qiov, qiov->niov); 27829ff7890SWen Congyang while (remaining_sectors > 0) { 27951b0a488SEric Blake int64_t count; 28051b0a488SEric Blake 281170d3bd3SAndrey Shinkevich ret = bdrv_is_allocated_above(top->bs, base->bs, false, 28251b0a488SEric Blake sector_num * BDRV_SECTOR_SIZE, 28351b0a488SEric Blake remaining_sectors * BDRV_SECTOR_SIZE, 28451b0a488SEric Blake &count); 28529ff7890SWen Congyang if (ret < 0) { 28629ff7890SWen Congyang goto out1; 28729ff7890SWen Congyang } 28829ff7890SWen Congyang 28951b0a488SEric Blake assert(QEMU_IS_ALIGNED(count, BDRV_SECTOR_SIZE)); 29051b0a488SEric Blake n = count >> BDRV_SECTOR_BITS; 29129ff7890SWen Congyang qemu_iovec_reset(&hd_qiov); 29251b0a488SEric Blake qemu_iovec_concat(&hd_qiov, qiov, bytes_done, count); 29329ff7890SWen Congyang 29429ff7890SWen Congyang target = ret ? top : base; 29504a11d87SEric Blake ret = bdrv_co_pwritev(target, sector_num * BDRV_SECTOR_SIZE, 29604a11d87SEric Blake n * BDRV_SECTOR_SIZE, &hd_qiov, 0); 29729ff7890SWen Congyang if (ret < 0) { 29829ff7890SWen Congyang goto out1; 29929ff7890SWen Congyang } 30029ff7890SWen Congyang 30129ff7890SWen Congyang remaining_sectors -= n; 30229ff7890SWen Congyang sector_num += n; 30351b0a488SEric Blake bytes_done += count; 30429ff7890SWen Congyang } 30529ff7890SWen Congyang 30629ff7890SWen Congyang out1: 30729ff7890SWen Congyang qemu_iovec_destroy(&hd_qiov); 30829ff7890SWen Congyang out: 30929ff7890SWen Congyang return ret; 31029ff7890SWen Congyang } 31129ff7890SWen Congyang 31229ff7890SWen Congyang static void secondary_do_checkpoint(BDRVReplicationState *s, Error **errp) 31329ff7890SWen Congyang { 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 328d470ad42SMax Reitz if (!s->active_disk->bs->drv) { 329d470ad42SMax Reitz error_setg(errp, "Active disk %s is ejected", 330d470ad42SMax Reitz s->active_disk->bs->node_name); 331d470ad42SMax Reitz return; 332d470ad42SMax Reitz } 333d470ad42SMax Reitz 334f844ec01SMax Reitz ret = bdrv_make_empty(s->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 3456ecbc6c5SKevin Wolf BlockBackend *blk = blk_new(qemu_get_current_aio_context(), 3466ecbc6c5SKevin Wolf BLK_PERM_WRITE, BLK_PERM_ALL); 3476ecbc6c5SKevin Wolf blk_insert_bs(blk, s->hidden_disk->bs, &local_err); 3486ecbc6c5SKevin Wolf if (local_err) { 3496ecbc6c5SKevin Wolf error_propagate(errp, local_err); 3506ecbc6c5SKevin Wolf blk_unref(blk); 3516ecbc6c5SKevin Wolf return; 3526ecbc6c5SKevin Wolf } 3536ecbc6c5SKevin Wolf 3546ecbc6c5SKevin Wolf ret = blk_make_empty(blk, errp); 3556ecbc6c5SKevin Wolf blk_unref(blk); 35629ff7890SWen Congyang if (ret < 0) { 35729ff7890SWen Congyang return; 35829ff7890SWen Congyang } 35929ff7890SWen Congyang } 36029ff7890SWen Congyang 3613c4e9647SAlberto Garcia /* This function is supposed to be called twice: 3623c4e9647SAlberto Garcia * first with writable = true, then with writable = false. 3633c4e9647SAlberto Garcia * The first call puts s->hidden_disk and s->secondary_disk in 3643c4e9647SAlberto Garcia * r/w mode, and the second puts them back in their original state. 3653c4e9647SAlberto Garcia */ 3668dd9006eSPaolo Bonzini static void reopen_backing_file(BlockDriverState *bs, bool writable, 36729ff7890SWen Congyang Error **errp) 36829ff7890SWen Congyang { 3698dd9006eSPaolo Bonzini BDRVReplicationState *s = bs->opaque; 37029ff7890SWen Congyang BlockReopenQueue *reopen_queue = NULL; 37129ff7890SWen Congyang Error *local_err = NULL; 37229ff7890SWen Congyang 37329ff7890SWen Congyang if (writable) { 3743c4e9647SAlberto Garcia s->orig_hidden_read_only = bdrv_is_read_only(s->hidden_disk->bs); 3753c4e9647SAlberto Garcia s->orig_secondary_read_only = bdrv_is_read_only(s->secondary_disk->bs); 37629ff7890SWen Congyang } 37729ff7890SWen Congyang 3781a63a907SKevin Wolf bdrv_subtree_drained_begin(s->hidden_disk->bs); 3791a63a907SKevin Wolf bdrv_subtree_drained_begin(s->secondary_disk->bs); 3801a63a907SKevin Wolf 3813c4e9647SAlberto Garcia if (s->orig_hidden_read_only) { 3823c4e9647SAlberto Garcia QDict *opts = qdict_new(); 3833c4e9647SAlberto Garcia qdict_put_bool(opts, BDRV_OPT_READ_ONLY, !writable); 3843c4e9647SAlberto Garcia reopen_queue = bdrv_reopen_queue(reopen_queue, s->hidden_disk->bs, 385077e8e20SAlberto Garcia opts, true); 38629ff7890SWen Congyang } 38729ff7890SWen Congyang 3883c4e9647SAlberto Garcia if (s->orig_secondary_read_only) { 3893c4e9647SAlberto Garcia QDict *opts = qdict_new(); 3903c4e9647SAlberto Garcia qdict_put_bool(opts, BDRV_OPT_READ_ONLY, !writable); 39129ff7890SWen Congyang reopen_queue = bdrv_reopen_queue(reopen_queue, s->secondary_disk->bs, 392077e8e20SAlberto Garcia opts, true); 39329ff7890SWen Congyang } 39429ff7890SWen Congyang 39529ff7890SWen Congyang if (reopen_queue) { 3965019aeceSAlberto Garcia bdrv_reopen_multiple(reopen_queue, &local_err); 39729ff7890SWen Congyang error_propagate(errp, local_err); 39829ff7890SWen Congyang } 3991a63a907SKevin Wolf 4001a63a907SKevin Wolf bdrv_subtree_drained_end(s->hidden_disk->bs); 4011a63a907SKevin Wolf bdrv_subtree_drained_end(s->secondary_disk->bs); 40229ff7890SWen Congyang } 40329ff7890SWen Congyang 4048dd9006eSPaolo Bonzini static void backup_job_cleanup(BlockDriverState *bs) 40529ff7890SWen Congyang { 4068dd9006eSPaolo Bonzini BDRVReplicationState *s = bs->opaque; 40729ff7890SWen Congyang BlockDriverState *top_bs; 40829ff7890SWen Congyang 409e140f4b7SLukas Straub s->backup_job = NULL; 410e140f4b7SLukas Straub 41129ff7890SWen Congyang top_bs = bdrv_lookup_bs(s->top_id, s->top_id, NULL); 41229ff7890SWen Congyang if (!top_bs) { 41329ff7890SWen Congyang return; 41429ff7890SWen Congyang } 41529ff7890SWen Congyang bdrv_op_unblock_all(top_bs, s->blocker); 41629ff7890SWen Congyang error_free(s->blocker); 4178dd9006eSPaolo Bonzini reopen_backing_file(bs, false, NULL); 41829ff7890SWen Congyang } 41929ff7890SWen Congyang 42029ff7890SWen Congyang static void backup_job_completed(void *opaque, int ret) 42129ff7890SWen Congyang { 4228dd9006eSPaolo Bonzini BlockDriverState *bs = opaque; 4238dd9006eSPaolo Bonzini BDRVReplicationState *s = bs->opaque; 42429ff7890SWen Congyang 4253c76c606SFam Zheng if (s->stage != BLOCK_REPLICATION_FAILOVER) { 42629ff7890SWen Congyang /* The backup job is cancelled unexpectedly */ 42729ff7890SWen Congyang s->error = -EIO; 42829ff7890SWen Congyang } 42929ff7890SWen Congyang 4308dd9006eSPaolo Bonzini backup_job_cleanup(bs); 43129ff7890SWen Congyang } 43229ff7890SWen Congyang 43329ff7890SWen Congyang static bool check_top_bs(BlockDriverState *top_bs, BlockDriverState *bs) 43429ff7890SWen Congyang { 43529ff7890SWen Congyang BdrvChild *child; 43629ff7890SWen Congyang 43729ff7890SWen Congyang /* The bs itself is the top_bs */ 43829ff7890SWen Congyang if (top_bs == bs) { 43929ff7890SWen Congyang return true; 44029ff7890SWen Congyang } 44129ff7890SWen Congyang 44229ff7890SWen Congyang /* Iterate over top_bs's children */ 44329ff7890SWen Congyang QLIST_FOREACH(child, &top_bs->children, next) { 44429ff7890SWen Congyang if (child->bs == bs || check_top_bs(child->bs, bs)) { 44529ff7890SWen Congyang return true; 44629ff7890SWen Congyang } 44729ff7890SWen Congyang } 44829ff7890SWen Congyang 44929ff7890SWen Congyang return false; 45029ff7890SWen Congyang } 45129ff7890SWen Congyang 45229ff7890SWen Congyang static void replication_start(ReplicationState *rs, ReplicationMode mode, 45329ff7890SWen Congyang Error **errp) 45429ff7890SWen Congyang { 45529ff7890SWen Congyang BlockDriverState *bs = rs->opaque; 45629ff7890SWen Congyang BDRVReplicationState *s; 45729ff7890SWen Congyang BlockDriverState *top_bs; 45829ff7890SWen Congyang int64_t active_length, hidden_length, disk_length; 45929ff7890SWen Congyang AioContext *aio_context; 46029ff7890SWen Congyang Error *local_err = NULL; 46129ff7890SWen Congyang 46229ff7890SWen Congyang aio_context = bdrv_get_aio_context(bs); 46329ff7890SWen Congyang aio_context_acquire(aio_context); 46429ff7890SWen Congyang s = bs->opaque; 46529ff7890SWen Congyang 46608ddb4ebSLukas Straub if (s->stage == BLOCK_REPLICATION_DONE || 46708ddb4ebSLukas Straub s->stage == BLOCK_REPLICATION_FAILOVER) { 46808ddb4ebSLukas Straub /* 46908ddb4ebSLukas Straub * This case happens when a secondary is promoted to primary. 47008ddb4ebSLukas Straub * Ignore the request because the secondary side of replication 47108ddb4ebSLukas Straub * doesn't have to do anything anymore. 47208ddb4ebSLukas Straub */ 47308ddb4ebSLukas Straub aio_context_release(aio_context); 47408ddb4ebSLukas Straub return; 47508ddb4ebSLukas Straub } 47608ddb4ebSLukas Straub 4773c76c606SFam Zheng if (s->stage != BLOCK_REPLICATION_NONE) { 47829ff7890SWen Congyang error_setg(errp, "Block replication is running or done"); 47929ff7890SWen Congyang aio_context_release(aio_context); 48029ff7890SWen Congyang return; 48129ff7890SWen Congyang } 48229ff7890SWen Congyang 48329ff7890SWen Congyang if (s->mode != mode) { 48429ff7890SWen Congyang error_setg(errp, "The parameter mode's value is invalid, needs %d," 48529ff7890SWen Congyang " but got %d", s->mode, mode); 48629ff7890SWen Congyang aio_context_release(aio_context); 48729ff7890SWen Congyang return; 48829ff7890SWen Congyang } 48929ff7890SWen Congyang 49029ff7890SWen Congyang switch (s->mode) { 49129ff7890SWen Congyang case REPLICATION_MODE_PRIMARY: 49229ff7890SWen Congyang break; 49329ff7890SWen Congyang case REPLICATION_MODE_SECONDARY: 49429ff7890SWen Congyang s->active_disk = bs->file; 49529ff7890SWen Congyang if (!s->active_disk || !s->active_disk->bs || 49629ff7890SWen Congyang !s->active_disk->bs->backing) { 49729ff7890SWen Congyang error_setg(errp, "Active disk doesn't have backing file"); 49829ff7890SWen Congyang aio_context_release(aio_context); 49929ff7890SWen Congyang return; 50029ff7890SWen Congyang } 50129ff7890SWen Congyang 50229ff7890SWen Congyang s->hidden_disk = s->active_disk->bs->backing; 50329ff7890SWen Congyang if (!s->hidden_disk->bs || !s->hidden_disk->bs->backing) { 50429ff7890SWen Congyang error_setg(errp, "Hidden disk doesn't have backing file"); 50529ff7890SWen Congyang aio_context_release(aio_context); 50629ff7890SWen Congyang return; 50729ff7890SWen Congyang } 50829ff7890SWen Congyang 50929ff7890SWen Congyang s->secondary_disk = s->hidden_disk->bs->backing; 51029ff7890SWen Congyang if (!s->secondary_disk->bs || !bdrv_has_blk(s->secondary_disk->bs)) { 51129ff7890SWen Congyang error_setg(errp, "The secondary disk doesn't have block backend"); 51229ff7890SWen Congyang aio_context_release(aio_context); 51329ff7890SWen Congyang return; 51429ff7890SWen Congyang } 51529ff7890SWen Congyang 51629ff7890SWen Congyang /* verify the length */ 51729ff7890SWen Congyang active_length = bdrv_getlength(s->active_disk->bs); 51829ff7890SWen Congyang hidden_length = bdrv_getlength(s->hidden_disk->bs); 51929ff7890SWen Congyang disk_length = bdrv_getlength(s->secondary_disk->bs); 52029ff7890SWen Congyang if (active_length < 0 || hidden_length < 0 || disk_length < 0 || 52129ff7890SWen Congyang active_length != hidden_length || hidden_length != disk_length) { 52229ff7890SWen Congyang error_setg(errp, "Active disk, hidden disk, secondary disk's length" 52329ff7890SWen Congyang " are not the same"); 52429ff7890SWen Congyang aio_context_release(aio_context); 52529ff7890SWen Congyang return; 52629ff7890SWen Congyang } 52729ff7890SWen Congyang 528d470ad42SMax Reitz /* Must be true, or the bdrv_getlength() calls would have failed */ 529d470ad42SMax Reitz assert(s->active_disk->bs->drv && s->hidden_disk->bs->drv); 530d470ad42SMax Reitz 53129ff7890SWen Congyang if (!s->active_disk->bs->drv->bdrv_make_empty || 53229ff7890SWen Congyang !s->hidden_disk->bs->drv->bdrv_make_empty) { 53329ff7890SWen Congyang error_setg(errp, 53429ff7890SWen Congyang "Active disk or hidden disk doesn't support make_empty"); 53529ff7890SWen Congyang aio_context_release(aio_context); 53629ff7890SWen Congyang return; 53729ff7890SWen Congyang } 53829ff7890SWen Congyang 53929ff7890SWen Congyang /* reopen the backing file in r/w mode */ 5408dd9006eSPaolo Bonzini reopen_backing_file(bs, true, &local_err); 54129ff7890SWen Congyang if (local_err) { 54229ff7890SWen Congyang error_propagate(errp, local_err); 54329ff7890SWen Congyang aio_context_release(aio_context); 54429ff7890SWen Congyang return; 54529ff7890SWen Congyang } 54629ff7890SWen Congyang 54729ff7890SWen Congyang /* start backup job now */ 54829ff7890SWen Congyang error_setg(&s->blocker, 54929ff7890SWen Congyang "Block device is in use by internal backup job"); 55029ff7890SWen Congyang 55129ff7890SWen Congyang top_bs = bdrv_lookup_bs(s->top_id, s->top_id, NULL); 55229ff7890SWen Congyang if (!top_bs || !bdrv_is_root_node(top_bs) || 55329ff7890SWen Congyang !check_top_bs(top_bs, bs)) { 55429ff7890SWen Congyang error_setg(errp, "No top_bs or it is invalid"); 5558dd9006eSPaolo Bonzini reopen_backing_file(bs, false, NULL); 55629ff7890SWen Congyang aio_context_release(aio_context); 55729ff7890SWen Congyang return; 55829ff7890SWen Congyang } 55929ff7890SWen Congyang bdrv_op_block_all(top_bs, s->blocker); 56029ff7890SWen Congyang bdrv_op_unblock(top_bs, BLOCK_OP_TYPE_DATAPLANE, s->blocker); 56129ff7890SWen Congyang 562cc19f177SVladimir Sementsov-Ogievskiy s->backup_job = backup_job_create( 563cc19f177SVladimir Sementsov-Ogievskiy NULL, s->secondary_disk->bs, s->hidden_disk->bs, 56400e30f05SVladimir Sementsov-Ogievskiy 0, MIRROR_SYNC_MODE_NONE, NULL, 0, false, NULL, 565111049a4SJohn Snow BLOCKDEV_ON_ERROR_REPORT, 566bb02b65cSKevin Wolf BLOCKDEV_ON_ERROR_REPORT, JOB_INTERNAL, 567111049a4SJohn Snow backup_job_completed, bs, NULL, &local_err); 56829ff7890SWen Congyang if (local_err) { 56929ff7890SWen Congyang error_propagate(errp, local_err); 5708dd9006eSPaolo Bonzini backup_job_cleanup(bs); 57129ff7890SWen Congyang aio_context_release(aio_context); 57229ff7890SWen Congyang return; 57329ff7890SWen Congyang } 574cc19f177SVladimir Sementsov-Ogievskiy job_start(&s->backup_job->job); 57529ff7890SWen Congyang break; 57629ff7890SWen Congyang default: 57729ff7890SWen Congyang aio_context_release(aio_context); 57829ff7890SWen Congyang abort(); 57929ff7890SWen Congyang } 58029ff7890SWen Congyang 5813c76c606SFam Zheng s->stage = BLOCK_REPLICATION_RUNNING; 58229ff7890SWen Congyang 58329ff7890SWen Congyang if (s->mode == REPLICATION_MODE_SECONDARY) { 58429ff7890SWen Congyang secondary_do_checkpoint(s, errp); 58529ff7890SWen Congyang } 58629ff7890SWen Congyang 58729ff7890SWen Congyang s->error = 0; 58829ff7890SWen Congyang aio_context_release(aio_context); 58929ff7890SWen Congyang } 59029ff7890SWen Congyang 59129ff7890SWen Congyang static void replication_do_checkpoint(ReplicationState *rs, Error **errp) 59229ff7890SWen Congyang { 59329ff7890SWen Congyang BlockDriverState *bs = rs->opaque; 59429ff7890SWen Congyang BDRVReplicationState *s; 59529ff7890SWen Congyang AioContext *aio_context; 59629ff7890SWen Congyang 59729ff7890SWen Congyang aio_context = bdrv_get_aio_context(bs); 59829ff7890SWen Congyang aio_context_acquire(aio_context); 59929ff7890SWen Congyang s = bs->opaque; 60029ff7890SWen Congyang 60108ddb4ebSLukas Straub if (s->stage == BLOCK_REPLICATION_DONE || 60208ddb4ebSLukas Straub s->stage == BLOCK_REPLICATION_FAILOVER) { 60308ddb4ebSLukas Straub /* 60408ddb4ebSLukas Straub * This case happens when a secondary was promoted to primary. 60508ddb4ebSLukas Straub * Ignore the request because the secondary side of replication 60608ddb4ebSLukas Straub * doesn't have to do anything anymore. 60708ddb4ebSLukas Straub */ 60808ddb4ebSLukas Straub aio_context_release(aio_context); 60908ddb4ebSLukas Straub return; 61008ddb4ebSLukas Straub } 61108ddb4ebSLukas Straub 61229ff7890SWen Congyang if (s->mode == REPLICATION_MODE_SECONDARY) { 61329ff7890SWen Congyang secondary_do_checkpoint(s, errp); 61429ff7890SWen Congyang } 61529ff7890SWen Congyang aio_context_release(aio_context); 61629ff7890SWen Congyang } 61729ff7890SWen Congyang 61829ff7890SWen Congyang static void replication_get_error(ReplicationState *rs, Error **errp) 61929ff7890SWen Congyang { 62029ff7890SWen Congyang BlockDriverState *bs = rs->opaque; 62129ff7890SWen Congyang BDRVReplicationState *s; 62229ff7890SWen Congyang AioContext *aio_context; 62329ff7890SWen Congyang 62429ff7890SWen Congyang aio_context = bdrv_get_aio_context(bs); 62529ff7890SWen Congyang aio_context_acquire(aio_context); 62629ff7890SWen Congyang s = bs->opaque; 62729ff7890SWen Congyang 62808ddb4ebSLukas Straub if (s->stage == BLOCK_REPLICATION_NONE) { 62929ff7890SWen Congyang error_setg(errp, "Block replication is not running"); 63029ff7890SWen Congyang aio_context_release(aio_context); 63129ff7890SWen Congyang return; 63229ff7890SWen Congyang } 63329ff7890SWen Congyang 63429ff7890SWen Congyang if (s->error) { 63529ff7890SWen Congyang error_setg(errp, "I/O error occurred"); 63629ff7890SWen Congyang aio_context_release(aio_context); 63729ff7890SWen Congyang return; 63829ff7890SWen Congyang } 63929ff7890SWen Congyang aio_context_release(aio_context); 64029ff7890SWen Congyang } 64129ff7890SWen Congyang 64229ff7890SWen Congyang static void replication_done(void *opaque, int ret) 64329ff7890SWen Congyang { 64429ff7890SWen Congyang BlockDriverState *bs = opaque; 64529ff7890SWen Congyang BDRVReplicationState *s = bs->opaque; 64629ff7890SWen Congyang 64729ff7890SWen Congyang if (ret == 0) { 6483c76c606SFam Zheng s->stage = BLOCK_REPLICATION_DONE; 64929ff7890SWen Congyang 65029ff7890SWen Congyang s->active_disk = NULL; 65129ff7890SWen Congyang s->secondary_disk = NULL; 65229ff7890SWen Congyang s->hidden_disk = NULL; 65329ff7890SWen Congyang s->error = 0; 65429ff7890SWen Congyang } else { 6553c76c606SFam Zheng s->stage = BLOCK_REPLICATION_FAILOVER_FAILED; 65629ff7890SWen Congyang s->error = -EIO; 65729ff7890SWen Congyang } 65829ff7890SWen Congyang } 65929ff7890SWen Congyang 66029ff7890SWen Congyang static void replication_stop(ReplicationState *rs, bool failover, Error **errp) 66129ff7890SWen Congyang { 66229ff7890SWen Congyang BlockDriverState *bs = rs->opaque; 66329ff7890SWen Congyang BDRVReplicationState *s; 66429ff7890SWen Congyang AioContext *aio_context; 66529ff7890SWen Congyang 66629ff7890SWen Congyang aio_context = bdrv_get_aio_context(bs); 66729ff7890SWen Congyang aio_context_acquire(aio_context); 66829ff7890SWen Congyang s = bs->opaque; 66929ff7890SWen Congyang 67008ddb4ebSLukas Straub if (s->stage == BLOCK_REPLICATION_DONE || 67108ddb4ebSLukas Straub s->stage == BLOCK_REPLICATION_FAILOVER) { 67208ddb4ebSLukas Straub /* 67308ddb4ebSLukas Straub * This case happens when a secondary was promoted to primary. 67408ddb4ebSLukas Straub * Ignore the request because the secondary side of replication 67508ddb4ebSLukas Straub * doesn't have to do anything anymore. 67608ddb4ebSLukas Straub */ 67708ddb4ebSLukas Straub aio_context_release(aio_context); 67808ddb4ebSLukas Straub return; 67908ddb4ebSLukas Straub } 68008ddb4ebSLukas Straub 6813c76c606SFam Zheng if (s->stage != BLOCK_REPLICATION_RUNNING) { 68229ff7890SWen Congyang error_setg(errp, "Block replication is not running"); 68329ff7890SWen Congyang aio_context_release(aio_context); 68429ff7890SWen Congyang return; 68529ff7890SWen Congyang } 68629ff7890SWen Congyang 68729ff7890SWen Congyang switch (s->mode) { 68829ff7890SWen Congyang case REPLICATION_MODE_PRIMARY: 6893c76c606SFam Zheng s->stage = BLOCK_REPLICATION_DONE; 69029ff7890SWen Congyang s->error = 0; 69129ff7890SWen Congyang break; 69229ff7890SWen Congyang case REPLICATION_MODE_SECONDARY: 69329ff7890SWen Congyang /* 69429ff7890SWen Congyang * This BDS will be closed, and the job should be completed 69529ff7890SWen Congyang * before the BDS is closed, because we will access hidden 69629ff7890SWen Congyang * disk, secondary disk in backup_job_completed(). 69729ff7890SWen Congyang */ 698cc19f177SVladimir Sementsov-Ogievskiy if (s->backup_job) { 699cc19f177SVladimir Sementsov-Ogievskiy job_cancel_sync(&s->backup_job->job); 70029ff7890SWen Congyang } 70129ff7890SWen Congyang 70229ff7890SWen Congyang if (!failover) { 70329ff7890SWen Congyang secondary_do_checkpoint(s, errp); 7043c76c606SFam Zheng s->stage = BLOCK_REPLICATION_DONE; 70529ff7890SWen Congyang aio_context_release(aio_context); 70629ff7890SWen Congyang return; 70729ff7890SWen Congyang } 70829ff7890SWen Congyang 7093c76c606SFam Zheng s->stage = BLOCK_REPLICATION_FAILOVER; 710cc19f177SVladimir Sementsov-Ogievskiy s->commit_job = commit_active_start( 711cc19f177SVladimir Sementsov-Ogievskiy NULL, s->active_disk->bs, s->secondary_disk->bs, 712bb02b65cSKevin Wolf JOB_INTERNAL, 0, BLOCKDEV_ON_ERROR_REPORT, 71378bbd910SFam Zheng NULL, replication_done, bs, true, errp); 71429ff7890SWen Congyang break; 71529ff7890SWen Congyang default: 71629ff7890SWen Congyang aio_context_release(aio_context); 71729ff7890SWen Congyang abort(); 71829ff7890SWen Congyang } 71929ff7890SWen Congyang aio_context_release(aio_context); 72029ff7890SWen Congyang } 72129ff7890SWen Congyang 7222654267cSMax Reitz static const char *const replication_strong_runtime_opts[] = { 7232654267cSMax Reitz REPLICATION_MODE, 7242654267cSMax Reitz REPLICATION_TOP_ID, 7252654267cSMax Reitz 7262654267cSMax Reitz NULL 7272654267cSMax Reitz }; 7282654267cSMax Reitz 729782b9d06SAlberto Garcia static BlockDriver bdrv_replication = { 73029ff7890SWen Congyang .format_name = "replication", 73129ff7890SWen Congyang .instance_size = sizeof(BDRVReplicationState), 73229ff7890SWen Congyang 73329ff7890SWen Congyang .bdrv_open = replication_open, 73429ff7890SWen Congyang .bdrv_close = replication_close, 73537a9051cSChanglong Xie .bdrv_child_perm = replication_child_perm, 73629ff7890SWen Congyang 73729ff7890SWen Congyang .bdrv_getlength = replication_getlength, 73829ff7890SWen Congyang .bdrv_co_readv = replication_co_readv, 73929ff7890SWen Congyang .bdrv_co_writev = replication_co_writev, 74029ff7890SWen Congyang 74129ff7890SWen Congyang .is_filter = true, 74229ff7890SWen Congyang 74329ff7890SWen Congyang .has_variable_length = true, 7442654267cSMax Reitz .strong_runtime_opts = replication_strong_runtime_opts, 74529ff7890SWen Congyang }; 74629ff7890SWen Congyang 74729ff7890SWen Congyang static void bdrv_replication_init(void) 74829ff7890SWen Congyang { 74929ff7890SWen Congyang bdrv_register(&bdrv_replication); 75029ff7890SWen Congyang } 75129ff7890SWen Congyang 75229ff7890SWen Congyang block_init(bdrv_replication_init); 753