savevm-async: iothreads fixups

[pve-qemu-kvm.git] / debian / patches / pve / 0046-convert-savevm-async-to-threads.patch

From f255c2a02017dbefc3bcc786762b1de3c5aa5c89 Mon Sep 17 00:00:00 2001
From: Wolfgang Bumiller <w.bumiller@proxmox.com>
Date: Tue, 8 Nov 2016 11:13:06 +0100
Subject: [PATCH 46/46] convert savevm-async to threads

---
 savevm-async.c | 143 +++++++++++++++++++++++++++++++++++----------------------
 1 file changed, 88 insertions(+), 55 deletions(-)

diff --git a/savevm-async.c b/savevm-async.c
index 05b5b19..c4afc3b 100644
--- a/savevm-async.c
+++ b/savevm-async.c
@@ -48,6 +48,8 @@ static struct SnapshotState {
     int saved_vm_running;
     QEMUFile *file;
     int64_t total_time;
+    QEMUBH *cleanup_bh;
+    QemuThread thread;
 } snap_state;
 
 SaveVMInfo *qmp_query_savevm(Error **errp)
@@ -135,19 +137,6 @@ static void save_snapshot_error(const char *fmt, ...)
     g_free (msg);
 
     snap_state.state = SAVE_STATE_ERROR;
-
-    save_snapshot_cleanup();
-}
-
-static void save_snapshot_completed(void)
-{
-    DPRINTF("save_snapshot_completed\n");
-
-    if (save_snapshot_cleanup() < 0) {
-        snap_state.state = SAVE_STATE_ERROR;
-    } else {
-        snap_state.state = SAVE_STATE_COMPLETED;
-    }
 }
 
 static int block_state_close(void *opaque)
@@ -156,36 +145,76 @@ static int block_state_close(void *opaque)
     return blk_flush(snap_state.target);
 }
 
+typedef struct BlkRwCo {
+    int64_t offset;
+    QEMUIOVector *qiov;
+    int ret;
+} BlkRwCo;
+
+static void block_state_write_entry(void *opaque) {
+    BlkRwCo *rwco = opaque;
+    rwco->ret = blk_co_pwritev(snap_state.target, rwco->offset, rwco->qiov->size,
+                               rwco->qiov, 0);
+}
+
 static ssize_t block_state_writev_buffer(void *opaque, struct iovec *iov,
                                          int iovcnt, int64_t pos)
 {
-    int ret;
     QEMUIOVector qiov;
+    AioContext *aio_context;
+    Coroutine *co;
+    BlkRwCo rwco;
+
+    assert(pos == snap_state.bs_pos);
+    rwco = (BlkRwCo) {
+        .offset = pos,
+        .qiov = &qiov,
+        .ret = NOT_DONE,
+    };
 
     qemu_iovec_init_external(&qiov, iov, iovcnt);
-    ret = blk_co_pwritev(snap_state.target, pos, qiov.size, &qiov, 0);
-    if (ret < 0) {
-        return ret;
+
+    aio_context = blk_get_aio_context(snap_state.target);
+    aio_context_acquire(aio_context);
+    co = qemu_coroutine_create(&block_state_write_entry, &rwco);
+    qemu_coroutine_enter(co);
+    while (rwco.ret == NOT_DONE) {
+        aio_poll(aio_context, true);
     }
+    aio_context_release(aio_context);
+
     snap_state.bs_pos += qiov.size;
     return qiov.size;
 }
 
-static int store_and_stop(void) {
-    if (global_state_store()) {
-        save_snapshot_error("Error saving global state");
-        return 1;
+static void process_savevm_cleanup(void *opaque)
+{
+    int ret;
+    qemu_bh_delete(snap_state.cleanup_bh);
+    snap_state.cleanup_bh = NULL;
+    qemu_mutex_unlock_iothread();
+    qemu_thread_join(&snap_state.thread);
+    qemu_mutex_lock_iothread();
+    ret = save_snapshot_cleanup();
+    if (ret < 0) {
+        save_snapshot_error("save_snapshot_cleanup error %d", ret);
+    } else if (snap_state.state == SAVE_STATE_ACTIVE) {
+        snap_state.state = SAVE_STATE_COMPLETED;
+    } else {
+        save_snapshot_error("process_savevm_cleanup: invalid state: %d",
+                            snap_state.state);
     }
-    if (runstate_is_running()) {
-        vm_stop(RUN_STATE_SAVE_VM);
+    if (snap_state.saved_vm_running) {
+        vm_start();
+        snap_state.saved_vm_running = false;
     }
-    return 0;
 }
 
-static void process_savevm_co(void *opaque)
+static void *process_savevm_thread(void *opaque)
 {
     int ret;
     int64_t maxlen;
+
     MigrationParams params = {
         .blk = 0,
         .shared = 0
@@ -193,14 +222,15 @@ static void process_savevm_co(void *opaque)
 
     snap_state.state = SAVE_STATE_ACTIVE;
 
-    qemu_mutex_unlock_iothread();
+    rcu_register_thread();
+
     qemu_savevm_state_header(snap_state.file);
     ret = qemu_savevm_state_begin(snap_state.file, &params);
-    qemu_mutex_lock_iothread();
 
     if (ret < 0) {
         save_snapshot_error("qemu_savevm_state_begin failed");
-        return;
+        rcu_unregister_thread();
+        return NULL;
     }
 
     while (snap_state.state == SAVE_STATE_ACTIVE) {
@@ -209,41 +239,43 @@ static void process_savevm_co(void *opaque)
         qemu_savevm_state_pending(snap_state.file, 0, &pend_nonpost, &pend_post);
         pending_size = pend_post + pend_nonpost;
 
-        if (pending_size) {
-                ret = qemu_savevm_state_iterate(snap_state.file, false);
-                if (ret < 0) {
-                    save_snapshot_error("qemu_savevm_state_iterate error %d", ret);
-                    break;
-                }
-                DPRINTF("savevm inerate pending size %lu ret %d\n", pending_size, ret);
+        maxlen = blk_getlength(snap_state.target) - 30*1024*1024;
+
+        if (pending_size > 400000 && snap_state.bs_pos + pending_size < maxlen) {
+            qemu_mutex_lock_iothread();
+            ret = qemu_savevm_state_iterate(snap_state.file, false);
+            if (ret < 0) {
+                save_snapshot_error("qemu_savevm_state_iterate error %d", ret);
+                break;
+            }
+            qemu_mutex_unlock_iothread();
+            DPRINTF("savevm inerate pending size %lu ret %d\n", pending_size, ret);
         } else {
-            DPRINTF("done iterating\n");
-            if (store_and_stop())
+            qemu_mutex_lock_iothread();
+            qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
+            ret = global_state_store();
+            if (ret) {
+                save_snapshot_error("global_state_store error %d", ret);
+                break;
+            }
+            ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
+            if (ret < 0) {
+                save_snapshot_error("vm_stop_force_state error %d", ret);
                 break;
+            }
             DPRINTF("savevm inerate finished\n");
             qemu_savevm_state_complete_precopy(snap_state.file, false);
+            qemu_savevm_state_cleanup();
             DPRINTF("save complete\n");
-            save_snapshot_completed();
             break;
         }
-
-        /* stop the VM if we get to the end of available space,
-         * or if pending_size is just a few MB
-         */
-        maxlen = blk_getlength(snap_state.target) - 30*1024*1024;
-        if ((pending_size < 100000) ||
-            ((snap_state.bs_pos + pending_size) >= maxlen)) {
-            if (store_and_stop())
-                break;
-        }
     }
 
-    if(snap_state.state == SAVE_STATE_CANCELLED) {
-        save_snapshot_completed();
-        Error *errp = NULL;
-        qmp_savevm_end(&errp);
-    }
+    qemu_bh_schedule(snap_state.cleanup_bh);
+    qemu_mutex_unlock_iothread();
 
+    rcu_unregister_thread();
+    return NULL;
 }
 
 static const QEMUFileOps block_file_ops = {
@@ -306,8 +338,9 @@ void qmp_savevm_start(bool has_statefile, const char *statefile, Error **errp)
     error_setg(&snap_state.blocker, "block device is in use by savevm");
     blk_op_block_all(snap_state.target, snap_state.blocker);
 
-    Coroutine *co = qemu_coroutine_create(process_savevm_co, NULL);
-    qemu_coroutine_enter(co);
+    snap_state.cleanup_bh = qemu_bh_new(process_savevm_cleanup, &snap_state);
+    qemu_thread_create(&snap_state.thread, "savevm-async", process_savevm_thread,
+                       NULL, QEMU_THREAD_JOINABLE);
 
     return;
 
-- 
2.1.4