From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001 From: Stefan Reiter Date: Mon, 28 Sep 2020 13:40:51 +0200 Subject: [PATCH] PVE-Backup: Use more coroutines and don't block on finishing proxmox_backup_co_finish is already async, but previously we would wait for the coroutine using block_on_coroutine_fn(). Avoid this by scheduling pvebackup_co_complete_stream (and thus pvebackup_co_cleanup) as a real coroutine when calling from pvebackup_complete_cb. This is ok, since complete_stream uses the backup_mutex internally to synchronize, and other streams can happily continue writing in the meantime anyway. To accomodate, backup_mutex is converted to a CoMutex. This means converting every user to a coroutine. This is not just useful here, but will come in handy once this series[0] is merged, and QMP calls can be yield-able coroutines too. Then we can also finally get rid of block_on_coroutine_fn. Cases of aio_context_acquire/release from within what is now a coroutine are changed to aio_co_reschedule_self, which works since a running coroutine always holds the aio lock for the context it is running in. job_cancel_sync is called from a BH since it can't be run from a coroutine (uses AIO_WAIT_WHILE internally). Same thing for create_backup_jobs, which is converted to a BH too. To communicate the finishing state, a new property is introduced to query-backup: 'finishing'. A new state is explicitly not used, since that would break compatibility with older qemu-server versions. [0] https://lists.gnu.org/archive/html/qemu-devel/2020-09/msg03515.html Signed-off-by: Stefan Reiter --- pve-backup.c | 148 ++++++++++++++++++++++++++----------------- qapi/block-core.json | 5 +- 2 files changed, 95 insertions(+), 58 deletions(-) diff --git a/pve-backup.c b/pve-backup.c index f3df43eb8c..f3b8ce1f3a 100644 --- a/pve-backup.c +++ b/pve-backup.c @@ -33,7 +33,9 @@ const char *PBS_BITMAP_NAME = "pbs-incremental-dirty-bitmap"; static struct PVEBackupState { struct { - // Everithing accessed from qmp_backup_query command is protected using lock + // Everything accessed from qmp_backup_query command is protected using + // this lock. Do NOT hold this lock for long times, as it is sometimes + // acquired from coroutines, and thus any wait time may block the guest. QemuMutex lock; Error *error; time_t start_time; @@ -47,20 +49,21 @@ static struct PVEBackupState { size_t reused; size_t zero_bytes; GList *bitmap_list; + bool finishing; } stat; int64_t speed; VmaWriter *vmaw; ProxmoxBackupHandle *pbs; GList *di_list; JobTxn *txn; - QemuMutex backup_mutex; + CoMutex backup_mutex; CoMutex dump_callback_mutex; } backup_state; static void pvebackup_init(void) { qemu_mutex_init(&backup_state.stat.lock); - qemu_mutex_init(&backup_state.backup_mutex); + qemu_co_mutex_init(&backup_state.backup_mutex); qemu_co_mutex_init(&backup_state.dump_callback_mutex); } @@ -72,6 +75,7 @@ typedef struct PVEBackupDevInfo { size_t size; uint64_t block_size; uint8_t dev_id; + int completed_ret; // INT_MAX if not completed char targetfile[PATH_MAX]; BdrvDirtyBitmap *bitmap; BlockDriverState *target; @@ -227,12 +231,12 @@ pvebackup_co_dump_vma_cb( } // assumes the caller holds backup_mutex -static void coroutine_fn pvebackup_co_cleanup(void *unused) +static void coroutine_fn pvebackup_co_cleanup(void) { assert(qemu_in_coroutine()); qemu_mutex_lock(&backup_state.stat.lock); - backup_state.stat.end_time = time(NULL); + backup_state.stat.finishing = true; qemu_mutex_unlock(&backup_state.stat.lock); if (backup_state.vmaw) { @@ -261,12 +265,29 @@ static void coroutine_fn pvebackup_co_cleanup(void *unused) g_list_free(backup_state.di_list); backup_state.di_list = NULL; + + qemu_mutex_lock(&backup_state.stat.lock); + backup_state.stat.end_time = time(NULL); + backup_state.stat.finishing = false; + qemu_mutex_unlock(&backup_state.stat.lock); } -// assumes the caller holds backup_mutex -static void coroutine_fn pvebackup_complete_stream(void *opaque) +static void coroutine_fn pvebackup_co_complete_stream(void *opaque) { PVEBackupDevInfo *di = opaque; + int ret = di->completed_ret; + + qemu_co_mutex_lock(&backup_state.backup_mutex); + + if (ret < 0) { + Error *local_err = NULL; + error_setg(&local_err, "job failed with err %d - %s", ret, strerror(-ret)); + pvebackup_propagate_error(local_err); + } + + di->bs = NULL; + + assert(di->target == NULL); bool error_or_canceled = pvebackup_error_or_canceled(); @@ -281,27 +302,6 @@ static void coroutine_fn pvebackup_complete_stream(void *opaque) pvebackup_propagate_error(local_err); } } -} - -static void pvebackup_complete_cb(void *opaque, int ret) -{ - assert(!qemu_in_coroutine()); - - PVEBackupDevInfo *di = opaque; - - qemu_mutex_lock(&backup_state.backup_mutex); - - if (ret < 0) { - Error *local_err = NULL; - error_setg(&local_err, "job failed with err %d - %s", ret, strerror(-ret)); - pvebackup_propagate_error(local_err); - } - - di->bs = NULL; - - assert(di->target == NULL); - - block_on_coroutine_fn(pvebackup_complete_stream, di); // remove self from job list backup_state.di_list = g_list_remove(backup_state.di_list, di); @@ -310,21 +310,49 @@ static void pvebackup_complete_cb(void *opaque, int ret) /* call cleanup if we're the last job */ if (!g_list_first(backup_state.di_list)) { - block_on_coroutine_fn(pvebackup_co_cleanup, NULL); + pvebackup_co_cleanup(); } - qemu_mutex_unlock(&backup_state.backup_mutex); + qemu_co_mutex_unlock(&backup_state.backup_mutex); } -static void pvebackup_cancel(void) +static void pvebackup_complete_cb(void *opaque, int ret) { - assert(!qemu_in_coroutine()); + PVEBackupDevInfo *di = opaque; + di->completed_ret = ret; + + /* + * Schedule stream cleanup in async coroutine. close_image and finish might + * take a while, so we can't block on them here. This way it also doesn't + * matter if we're already running in a coroutine or not. + * Note: di is a pointer to an entry in the global backup_state struct, so + * it stays valid. + */ + Coroutine *co = qemu_coroutine_create(pvebackup_co_complete_stream, di); + aio_co_enter(qemu_get_aio_context(), co); +} +/* + * job_cancel(_sync) does not like to be called from coroutines, so defer to + * main loop processing via a bottom half. + */ +static void job_cancel_bh(void *opaque) { + CoCtxData *data = (CoCtxData*)opaque; + Job *job = (Job*)data->data; + AioContext *job_ctx = job->aio_context; + aio_context_acquire(job_ctx); + job_cancel_sync(job); + aio_context_release(job_ctx); + aio_co_enter(data->ctx, data->co); +} + +static void coroutine_fn pvebackup_co_cancel(void *opaque) +{ Error *cancel_err = NULL; error_setg(&cancel_err, "backup canceled"); pvebackup_propagate_error(cancel_err); - qemu_mutex_lock(&backup_state.backup_mutex); + qemu_co_mutex_lock(&backup_state.backup_mutex); if (backup_state.vmaw) { /* make sure vma writer does not block anymore */ @@ -342,27 +370,22 @@ static void pvebackup_cancel(void) ((PVEBackupDevInfo *)bdi->data)->job : NULL; - /* ref the job before releasing the mutex, just to be safe */ if (cancel_job) { - job_ref(&cancel_job->job); + CoCtxData data = { + .ctx = qemu_get_current_aio_context(), + .co = qemu_coroutine_self(), + .data = &cancel_job->job, + }; + aio_bh_schedule_oneshot(data.ctx, job_cancel_bh, &data); + qemu_coroutine_yield(); } - /* job_cancel_sync may enter the job, so we need to release the - * backup_mutex to avoid deadlock */ - qemu_mutex_unlock(&backup_state.backup_mutex); - - if (cancel_job) { - AioContext *aio_context = cancel_job->job.aio_context; - aio_context_acquire(aio_context); - job_cancel_sync(&cancel_job->job); - job_unref(&cancel_job->job); - aio_context_release(aio_context); - } + qemu_co_mutex_unlock(&backup_state.backup_mutex); } void qmp_backup_cancel(Error **errp) { - pvebackup_cancel(); + block_on_coroutine_fn(pvebackup_co_cancel, NULL); } // assumes the caller holds backup_mutex @@ -415,6 +438,14 @@ static int coroutine_fn pvebackup_co_add_config( goto out; } +/* + * backup_job_create can *not* be run from a coroutine (and requires an + * acquired AioContext), so this can't either. + * This does imply that this function cannot run with backup_mutex acquired. + * That is ok because it is only ever called between setting up the backup_state + * struct and starting the jobs, and from within a QMP call. This means that no + * other QMP call can interrupt, and no background job is running yet. + */ static bool create_backup_jobs(void) { assert(!qemu_in_coroutine()); @@ -523,11 +554,12 @@ typedef struct QmpBackupTask { UuidInfo *result; } QmpBackupTask; -// assumes the caller holds backup_mutex static void coroutine_fn pvebackup_co_prepare(void *opaque) { assert(qemu_in_coroutine()); + qemu_co_mutex_lock(&backup_state.backup_mutex); + QmpBackupTask *task = opaque; task->result = NULL; // just to be sure @@ -548,8 +580,9 @@ static void coroutine_fn pvebackup_co_prepare(void *opaque) const char *firewall_name = "qemu-server.fw"; if (backup_state.di_list) { - error_set(task->errp, ERROR_CLASS_GENERIC_ERROR, + error_set(task->errp, ERROR_CLASS_GENERIC_ERROR, "previous backup not finished"); + qemu_co_mutex_unlock(&backup_state.backup_mutex); return; } @@ -616,6 +649,8 @@ static void coroutine_fn pvebackup_co_prepare(void *opaque) } di->size = size; total += size; + + di->completed_ret = INT_MAX; } uuid_generate(uuid); @@ -847,6 +882,7 @@ static void coroutine_fn pvebackup_co_prepare(void *opaque) backup_state.stat.dirty = total - backup_state.stat.reused; backup_state.stat.transferred = 0; backup_state.stat.zero_bytes = 0; + backup_state.stat.finishing = false; qemu_mutex_unlock(&backup_state.stat.lock); @@ -861,6 +897,8 @@ static void coroutine_fn pvebackup_co_prepare(void *opaque) uuid_info->UUID = uuid_str; task->result = uuid_info; + + qemu_co_mutex_unlock(&backup_state.backup_mutex); return; err_mutex: @@ -903,6 +941,8 @@ err: } task->result = NULL; + + qemu_co_mutex_unlock(&backup_state.backup_mutex); return; } @@ -956,22 +996,15 @@ UuidInfo *qmp_backup( .errp = errp, }; - qemu_mutex_lock(&backup_state.backup_mutex); - block_on_coroutine_fn(pvebackup_co_prepare, &task); if (*errp == NULL) { bool errors = create_backup_jobs(); - qemu_mutex_unlock(&backup_state.backup_mutex); if (!errors) { - /* start the first job in the transaction - * note: this might directly enter the job, so we need to do this - * after unlocking the backup_mutex */ + // start the first job in the transaction job_txn_start_seq(backup_state.txn); } - } else { - qemu_mutex_unlock(&backup_state.backup_mutex); } return task.result; @@ -1025,6 +1058,7 @@ BackupStatus *qmp_query_backup(Error **errp) info->transferred = backup_state.stat.transferred; info->has_reused = true; info->reused = backup_state.stat.reused; + info->finishing = backup_state.stat.finishing; qemu_mutex_unlock(&backup_state.stat.lock); diff --git a/qapi/block-core.json b/qapi/block-core.json index b368371e8e..b90d6abe64 100644 --- a/qapi/block-core.json +++ b/qapi/block-core.json @@ -775,12 +775,15 @@ # # @uuid: uuid for this backup job # +# @finishing: if status='active' and finishing=true, then the backup process is +# waiting for the target to finish. +# ## { 'struct': 'BackupStatus', 'data': {'*status': 'str', '*errmsg': 'str', '*total': 'int', '*dirty': 'int', '*transferred': 'int', '*zero-bytes': 'int', '*reused': 'int', '*start-time': 'int', '*end-time': 'int', - '*backup-file': 'str', '*uuid': 'str' } } + '*backup-file': 'str', '*uuid': 'str', 'finishing': 'bool' } } ## # @BackupFormat: