debian/patches/pve/0011-convert-savevm-async-to-threads.patch

   1 From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
   2 From: Wolfgang Bumiller <w.bumiller@proxmox.com>
   3 Date: Tue, 8 Nov 2016 11:13:06 +0100
   4 Subject: [PATCH] convert savevm-async to threads
   5
   6 ---
   7  savevm-async.c | 143 +++++++++++++++++++++++++++++++++++----------------------
   8  1 file changed, 87 insertions(+), 56 deletions(-)
   9
  10 diff --git a/savevm-async.c b/savevm-async.c
  11 index 897134ab5a..96523c88ae 100644
  12 --- a/savevm-async.c
  13 +++ b/savevm-async.c
  14 @@ -43,6 +43,8 @@ static struct SnapshotState {
  15      int saved_vm_running;
  16      QEMUFile *file;
  17      int64_t total_time;
  18 +    QEMUBH *cleanup_bh;
  19 +    QemuThread thread;
  20  } snap_state;
  21
  22  SaveVMInfo *qmp_query_savevm(Error **errp)
  23 @@ -130,19 +132,6 @@ static void save_snapshot_error(const char *fmt, ...)
  24      g_free (msg);
  25
  26      snap_state.state = SAVE_STATE_ERROR;
  27 -
  28 -    save_snapshot_cleanup();
  29 -}
  30 -
  31 -static void save_snapshot_completed(void)
  32 -{
  33 -    DPRINTF("save_snapshot_completed\n");
  34 -
  35 -    if (save_snapshot_cleanup() < 0) {
  36 -        snap_state.state = SAVE_STATE_ERROR;
  37 -    } else {
  38 -        snap_state.state = SAVE_STATE_COMPLETED;
  39 -    }
  40  }
  41
  42  static int block_state_close(void *opaque)
  43 @@ -151,48 +140,86 @@ static int block_state_close(void *opaque)
  44      return blk_flush(snap_state.target);
  45  }
  46
  47 +typedef struct BlkRwCo {
  48 +    int64_t offset;
  49 +    QEMUIOVector *qiov;
  50 +    int ret;
  51 +} BlkRwCo;
  52 +
  53 +static void block_state_write_entry(void *opaque) {
  54 +    BlkRwCo *rwco = opaque;
  55 +    rwco->ret = blk_co_pwritev(snap_state.target, rwco->offset, rwco->qiov->size,
  56 +                               rwco->qiov, 0);
  57 +}
  58 +
  59  static ssize_t block_state_writev_buffer(void *opaque, struct iovec *iov,
  60                                           int iovcnt, int64_t pos)
  61  {
  62 -    int ret;
  63      QEMUIOVector qiov;
  64 +    AioContext *aio_context;
  65 +    Coroutine *co;
  66 +    BlkRwCo rwco;
  67 +
  68 +    assert(pos == snap_state.bs_pos);
  69 +    rwco = (BlkRwCo) {
  70 +        .offset = pos,
  71 +        .qiov = &qiov,
  72 +        .ret = NOT_DONE,
  73 +    };
  74
  75      qemu_iovec_init_external(&qiov, iov, iovcnt);
  76 -    ret = blk_co_pwritev(snap_state.target, pos, qiov.size, &qiov, 0);
  77 -    if (ret < 0) {
  78 -        return ret;
  79 +
  80 +    aio_context = blk_get_aio_context(snap_state.target);
  81 +    aio_context_acquire(aio_context);
  82 +    co = qemu_coroutine_create(&block_state_write_entry, &rwco);
  83 +    qemu_coroutine_enter(co);
  84 +    while (rwco.ret == NOT_DONE) {
  85 +        aio_poll(aio_context, true);
  86      }
  87 +    aio_context_release(aio_context);
  88 +
  89      snap_state.bs_pos += qiov.size;
  90      return qiov.size;
  91  }
  92
  93 -static int store_and_stop(void) {
  94 -    if (global_state_store()) {
  95 -        save_snapshot_error("Error saving global state");
  96 -        return 1;
  97 +static void process_savevm_cleanup(void *opaque)
  98 +{
  99 +    int ret;
 100 +    qemu_bh_delete(snap_state.cleanup_bh);
 101 +    snap_state.cleanup_bh = NULL;
 102 +    qemu_mutex_unlock_iothread();
 103 +    qemu_thread_join(&snap_state.thread);
 104 +    qemu_mutex_lock_iothread();
 105 +    ret = save_snapshot_cleanup();
 106 +    if (ret < 0) {
 107 +        save_snapshot_error("save_snapshot_cleanup error %d", ret);
 108 +    } else if (snap_state.state == SAVE_STATE_ACTIVE) {
 109 +        snap_state.state = SAVE_STATE_COMPLETED;
 110 +    } else {
 111 +        save_snapshot_error("process_savevm_cleanup: invalid state: %d",
 112 +                            snap_state.state);
 113      }
 114 -    if (runstate_is_running()) {
 115 -        vm_stop(RUN_STATE_SAVE_VM);
 116 +    if (snap_state.saved_vm_running) {
 117 +        vm_start();
 118 +        snap_state.saved_vm_running = false;
 119      }
 120 -    return 0;
 121  }
 122
 123 -static void process_savevm_co(void *opaque)
 124 +static void *process_savevm_thread(void *opaque)
 125  {
 126      int ret;
 127      int64_t maxlen;
 128
 129 -    snap_state.state = SAVE_STATE_ACTIVE;
 130 +    rcu_register_thread();
 131
 132 -    qemu_mutex_unlock_iothread();
 133      qemu_savevm_state_header(snap_state.file);
 134      qemu_savevm_state_setup(snap_state.file);
 135      ret = qemu_file_get_error(snap_state.file);
 136 -    qemu_mutex_lock_iothread();
 137
 138      if (ret < 0) {
 139          save_snapshot_error("qemu_savevm_state_setup failed");
 140 -        return;
 141 +        rcu_unregister_thread();
 142 +        return NULL;
 143      }
 144
 145      while (snap_state.state == SAVE_STATE_ACTIVE) {
 146 @@ -201,17 +228,30 @@ static void process_savevm_co(void *opaque)
 147          qemu_savevm_state_pending(snap_state.file, 0, &pend_nonpost, &pend_post);
 148          pending_size = pend_post + pend_nonpost;
 149
 150 -        if (pending_size) {
 151 -                ret = qemu_savevm_state_iterate(snap_state.file, false);
 152 -                if (ret < 0) {
 153 -                    save_snapshot_error("qemu_savevm_state_iterate error %d", ret);
 154 -                    break;
 155 -                }
 156 -                DPRINTF("savevm inerate pending size %lu ret %d\n", pending_size, ret);
 157 +        maxlen = blk_getlength(snap_state.target) - 30*1024*1024;
 158 +
 159 +        if (pending_size > 400000 && snap_state.bs_pos + pending_size < maxlen) {
 160 +            qemu_mutex_lock_iothread();
 161 +            ret = qemu_savevm_state_iterate(snap_state.file, false);
 162 +            if (ret < 0) {
 163 +                save_snapshot_error("qemu_savevm_state_iterate error %d", ret);
 164 +                break;
 165 +            }
 166 +            qemu_mutex_unlock_iothread();
 167 +            DPRINTF("savevm inerate pending size %lu ret %d\n", pending_size, ret);
 168          } else {
 169 -            DPRINTF("done iterating\n");
 170 -            if (store_and_stop())
 171 +            qemu_mutex_lock_iothread();
 172 +            qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER);
 173 +            ret = global_state_store();
 174 +            if (ret) {
 175 +                save_snapshot_error("global_state_store error %d", ret);
 176                  break;
 177 +            }
 178 +            ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
 179 +            if (ret < 0) {
 180 +                save_snapshot_error("vm_stop_force_state error %d", ret);
 181 +                break;
 182 +            }
 183              DPRINTF("savevm inerate finished\n");
 184              /* upstream made the return value here inconsistent
 185               * (-1 instead of 'ret' in one case and 0 after flush which can
 186 @@ -223,28 +263,17 @@ static void process_savevm_co(void *opaque)
 187                      save_snapshot_error("qemu_savevm_state_iterate error %d", ret);
 188                      break;
 189              }
 190 +            qemu_savevm_state_cleanup();
 191              DPRINTF("save complete\n");
 192 -            save_snapshot_completed();
 193              break;
 194          }
 195 -
 196 -        /* stop the VM if we get to the end of available space,
 197 -         * or if pending_size is just a few MB
 198 -         */
 199 -        maxlen = blk_getlength(snap_state.target) - 30*1024*1024;
 200 -        if ((pending_size < 100000) ||
 201 -            ((snap_state.bs_pos + pending_size) >= maxlen)) {
 202 -            if (store_and_stop())
 203 -                break;
 204 -        }
 205      }
 206
 207 -    if(snap_state.state == SAVE_STATE_CANCELLED) {
 208 -        save_snapshot_completed();
 209 -        Error *errp = NULL;
 210 -        qmp_savevm_end(&errp);
 211 -    }
 212 +    qemu_bh_schedule(snap_state.cleanup_bh);
 213 +    qemu_mutex_unlock_iothread();
 214
 215 +    rcu_unregister_thread();
 216 +    return NULL;
 217  }
 218
 219  static const QEMUFileOps block_file_ops = {
 220 @@ -307,8 +336,10 @@ void qmp_savevm_start(bool has_statefile, const char *statefile, Error **errp)
 221      error_setg(&snap_state.blocker, "block device is in use by savevm");
 222      blk_op_block_all(snap_state.target, snap_state.blocker);
 223
 224 -    Coroutine *co = qemu_coroutine_create(process_savevm_co, NULL);
 225 -    qemu_coroutine_enter(co);
 226 +    snap_state.state = SAVE_STATE_ACTIVE;
 227 +    snap_state.cleanup_bh = qemu_bh_new(process_savevm_cleanup, &snap_state);
 228 +    qemu_thread_create(&snap_state.thread, "savevm-async", process_savevm_thread,
 229 +                       NULL, QEMU_THREAD_JOINABLE);
 230
 231      return;
 232
 233 --
 234 2.11.0
 235