On Mon, Jul 08, 2019 at 11:55:51 -0500, Eric Blake wrote:
Time to actually issue the QMP transactions that start and stop backup commands (for now, just pull mode, not push). Starting a job has to kick off several pre-req steps, then a transaction, and additionally spawn an NBD server for pull mode; ending a job as well as failing partway through beginning a job has to unwind the earlier steps. Implementing push mode, as well as incremental pull and checkpoint creation, is deferred to later patches.
Signed-off-by: Eric Blake <eblake@redhat.com> --- src/qemu/qemu_domain.c | 18 +- src/qemu/qemu_driver.c | 310 ++++++++++++++++++++++++++++++++++- src/qemu/qemu_monitor_json.c | 4 + src/qemu/qemu_process.c | 9 + 4 files changed, 327 insertions(+), 14 deletions(-)
diff --git a/src/qemu/qemu_domain.c b/src/qemu/qemu_domain.c index 64c15d8ae6..70fe250880 100644 --- a/src/qemu/qemu_domain.c +++ b/src/qemu/qemu_domain.c @@ -2676,17 +2676,25 @@ qemuDomainObjPrivateXMLParseBlockjobs(virQEMUDriverPtr driver, { xmlNodePtr node; int tmp; + size_t i; VIR_AUTOFREE(char *) active = NULL;
if ((active = virXPathString("string(./blockjobs/@active)", ctxt)) && (tmp = virTristateBoolTypeFromString(active)) > 0) priv->reconnectBlockjobs = tmp;
- if ((node = virXPathNode("./domainbackup", ctxt)) && - !(priv->backup = virDomainBackupDefParseNode(ctxt->doc, node, - driver->xmlopt, - VIR_DOMAIN_BACKUP_PARSE_INTERNAL))) - return -1; + if ((node = virXPathNode("./domainbackup", ctxt))) { + if (!(priv->backup = virDomainBackupDefParseNode(ctxt->doc, node, + driver->xmlopt, + VIR_DOMAIN_BACKUP_PARSE_INTERNAL))) + return -1; + /* The backup job is only stored in XML if backupBegin + * succeeded at exporting the disk, so no need to store disk + * state when we can just force-reset it to a known-good + * value. */ + for (i = 0; i < priv->backup->ndisks; i++) + priv->backup->disks[i].state = VIR_DOMAIN_BACKUP_DISK_STATE_EXPORT;
Even for push mode?
+ }
return 0; } diff --git a/src/qemu/qemu_driver.c b/src/qemu/qemu_driver.c index ec5f37fe94..d9abcfa4c8 100644 --- a/src/qemu/qemu_driver.c +++ b/src/qemu/qemu_driver.c @@ -17555,8 +17555,80 @@ qemuDomainCheckpointDelete(virDomainCheckpointPtr checkpoint, return ret; }
-static int qemuDomainBackupBegin(virDomainPtr domain, const char *diskXml, - const char *checkpointXml, unsigned int flags) +static int +qemuDomainBackupPrepare(virQEMUDriverPtr driver, virDomainObjPtr vm, + virDomainBackupDefPtr def) +{ + int ret = -1; + size_t i; + + if (qemuBlockNodeNamesDetect(driver, vm, QEMU_ASYNC_JOB_NONE) < 0)
See previous reviews.
+ goto cleanup; + for (i = 0; i < def->ndisks; i++) { + virDomainBackupDiskDef *disk = &def->disks[i]; + virStorageSourcePtr src = vm->def->disks[disk->idx]->src; + + if (!disk->store) + continue; + if (virAsprintf(&disk->store->nodeformat, "tmp-%s", disk->name) < 0) + goto cleanup; + if (!disk->store->format) + disk->store->format = VIR_STORAGE_FILE_QCOW2; + if (def->incremental) { + if (src->format != VIR_STORAGE_FILE_QCOW2) { + virReportError(VIR_ERR_OPERATION_UNSUPPORTED, + _("incremental backup of %s requires qcow2"), + disk->name); + goto cleanup; + } + } + } + ret = 0; + cleanup: + return ret; +} + +/* Called while monitor lock is held. Best-effort cleanup. */ +static int +qemuDomainBackupDiskCleanup(virQEMUDriverPtr driver, virDomainObjPtr vm, + virDomainBackupDiskDef *disk, bool incremental) +{ + qemuDomainObjPrivatePtr priv = vm->privateData; + const char *node = vm->def->disks[disk->idx]->src->nodeformat; + int ret = 0; + + if (!disk->store) + return 0; + if (disk->state >= VIR_DOMAIN_BACKUP_DISK_STATE_EXPORT) { + /* No real need to use nbd-server-remove, since we will + * shortly be calling nbd-server-stop. */ + } + if (incremental && disk->state >= VIR_DOMAIN_BACKUP_DISK_STATE_BITMAP && + qemuMonitorDeleteBitmap(priv->mon, node, disk->store->nodeformat) < 0) { + VIR_WARN("Unable to remove temp bitmap for disk %s after backup", + disk->name); + ret = -1; + } + if (disk->state >= VIR_DOMAIN_BACKUP_DISK_STATE_READY && + qemuMonitorBlockdevDel(priv->mon, disk->store->nodeformat) < 0) { + VIR_WARN("Unable to remove temp disk %s after backup", + disk->name); + ret = -1; + } + if (disk->state >= VIR_DOMAIN_BACKUP_DISK_STATE_LABEL) + qemuDomainStorageSourceAccessRevoke(driver, vm, disk->store);
You must not call this with @vm unlocked.
+ if (disk->state >= VIR_DOMAIN_BACKUP_DISK_STATE_CREATED && + disk->store->detected && unlink(disk->store->path) < 0) { + VIR_WARN("Unable to unlink temp disk %s after backup", + disk->store->path); + ret = -1; + } + return ret; +} + +static int +qemuDomainBackupBegin(virDomainPtr domain, const char *diskXml, + const char *checkpointXml, unsigned int flags) { virQEMUDriverPtr driver = domain->conn->privateData; virDomainObjPtr vm = NULL;
[....]
@@ -17609,25 +17688,145 @@ static int qemuDomainBackupBegin(virDomainPtr domain, const char *diskXml,
[...]
- if (virDomainBackupAlignDisks(def, vm->def, suffix) < 0) + if (virDomainBackupAlignDisks(def, vm->def, suffix) < 0 || + qemuDomainBackupPrepare(driver, vm, def) < 0) goto endjob;
/* actually start the checkpoint. 2x2 array of push/pull, full/incr, plus additional tweak if checkpoint requested */ - /* TODO: issue QMP commands: - - pull: nbd-server-start with <server> from user (or autogenerate server) - - push/pull: blockdev-add per <disk> + qemuDomainObjEnterMonitor(driver, vm); + /* - push/pull: blockdev-add per <disk> */ + for (i = 0; i < def->ndisks; i++) { + virDomainBackupDiskDef *disk = &def->disks[i]; + virJSONValuePtr file; + virStorageSourcePtr src = vm->def->disks[disk->idx]->src; + const char *node = src->nodeformat; + + if (!disk->store) + continue; + if (qemuDomainStorageFileInit(driver, vm, disk->store, src) < 0)
This takes @vm which is unlocked
+ goto endmon; + if (disk->store->detected) { + if (virStorageFileCreate(disk->store) < 0) { + virReportSystemError(errno, + _("failed to create image file '%s'"), + NULLSTR(disk->store->path)); + goto endmon; + } + disk->state = VIR_DOMAIN_BACKUP_DISK_STATE_CREATED; + } + if (qemuDomainStorageSourceAccessAllow(driver, vm, disk->store, false,
Also this requires @vm locked
+ true) < 0) + goto endmon; + disk->state = VIR_DOMAIN_BACKUP_DISK_STATE_LABEL; + if (disk->store->detected) { + virBuffer buf = VIR_BUFFER_INITIALIZER; + + /* Force initialization of scratch/target file to new qcow2 */ + if (!(cmd = virCommandNewArgList(qemuImgPath, + "create", + "-f", + virStorageFileFormatTypeToString(disk->store->format), + "-o", + NULL))) + goto endmon;
AFAIK if we hold the monitor lock too much (and this qualifies as a long-running operation) without calling actual monitor APIs the event loop will get stuck.
+ virBufferAsprintf(&buf, "backing_fmt=%s,backing_file=", + virStorageFileFormatTypeToString(src->format)); + virQEMUBuildBufferEscapeComma(&buf, src->path); + virCommandAddArgBuffer(cmd, &buf); + + virQEMUBuildBufferEscapeComma(&buf, disk->store->path); + virCommandAddArgBuffer(cmd, &buf); + if (virCommandRun(cmd, NULL) < 0) + goto endmon; + virCommandFree(cmd); + cmd = NULL; + } + + if (virJSONValueObjectCreate(&file, + "s:driver", "file", + "s:filename", disk->store->path, + NULL) < 0) + goto endmon; + if (virJSONValueObjectCreate(&json, + "s:driver", virStorageFileFormatTypeToString(disk->store->format), + "s:node-name", disk->store->nodeformat, + "a:file", &file, + "s:backing", node, NULL) < 0) {
Please use the designated blockdev-add JSON props generators. Also libvirt opted to open the 'file' separately with blockdev add so this should not do things differently.
+ virJSONValueFree(file); + goto endmon; + } + if (qemuMonitorBlockdevAdd(priv->mon, json) < 0) + goto endmon; + json = NULL; + disk->state = VIR_DOMAIN_BACKUP_DISK_STATE_READY; + } + + /* TODO: - incr: bitmap-add of tmp, bitmap-merge per <disk> - transaction, containing: - push+full: blockdev-backup sync:full @@ -17635,8 +17834,76 @@ static int qemuDomainBackupBegin(virDomainPtr domain, const char *diskXml, - pull+full: blockdev-backup sync:none - pull+incr: blockdev-backup sync:none, bitmap-disable of tmp - if checkpoint: bitmap-disable of old, bitmap-add of new + */ + if (!(json = virJSONValueNewArray())) + goto endmon; + for (i = 0; i < def->ndisks; i++) { + virDomainBackupDiskDef *disk = &def->disks[i]; + virStorageSourcePtr src = vm->def->disks[disk->idx]->src; + + if (!disk->store) + continue; + if (qemuMonitorJSONTransactionAdd(json, + "blockdev-backup", + "s:device", src->nodeformat,
Also since this does not seem to register the blockjob with the disk you don't get the interlocking with other blockjobs
+ "s:target", disk->store->nodeformat, + "s:sync", "none", + "s:job-id", disk->name,
This should be prefixed with job type.
+ NULL) < 0) + goto endmon;
[...]
@@ -17719,9 +17990,27 @@ static int qemuDomainBackupEnd(virDomainPtr domain, int id, unsigned int flags)
if (priv->backup->type != VIR_DOMAIN_BACKUP_TYPE_PUSH) want_abort = false; + def = priv->backup; + + /* We are going to modify the domain below. */ + if (qemuDomainObjBeginJob(driver, vm, QEMU_JOB_MODIFY) < 0) + goto cleanup; + + qemuDomainObjEnterMonitor(driver, vm); + if (def->type == VIR_DOMAIN_BACKUP_TYPE_PULL) + ret = qemuMonitorNBDServerStop(priv->mon); + for (i = 0; i < def->ndisks; i++) { + if (qemuMonitorBlockJobCancel(priv->mon,
If not all disks are part of the job def->disk[i].name might point to a disk which is not part of the backup and this would then attempt to cancel a job of that disk.
+ def->disks[i].name) < 0 || + qemuDomainBackupDiskCleanup(driver, vm, &def->disks[i], + !!def->incremental) < 0) + ret = -1; + } + if (qemuDomainObjExitMonitor(driver, vm) < 0 || ret < 0) { + ret = -1; + goto endjob; + }