+void MSG_vm_start(msg_vm_t vm)
+{
+ simcall_vm_start(vm);
+
+ TRACE_msg_vm_start(vm);
+}
+
+
+
+/** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
+ * @ingroup msg_VMs
+ *
+ * FIXME: No extra delay occurs. If you want to simulate this too, you want to
+ * use a #MSG_process_sleep() or something. I'm not quite sure.
+ */
+void MSG_vm_shutdown(msg_vm_t vm)
+{
+ /* msg_vm_t equals to msg_host_t */
+ simcall_vm_shutdown(vm);
+
+ // TRACE_msg_vm_(vm);
+}
+
+
+
+/* We have two mailboxes. mbox is used to transfer migration data between
+ * source and destination PMs. mbox_ctl is used to detect the completion of a
+ * migration. The names of these mailboxes must not conflict with others. */
+static inline char *get_mig_mbox_src_dst(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
+{
+ char *vm_name = sg_host_name(vm);
+ char *src_pm_name = sg_host_name(src_pm);
+ char *dst_pm_name = sg_host_name(dst_pm);
+
+ return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
+}
+
+static inline char *get_mig_mbox_ctl(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
+{
+ char *vm_name = sg_host_name(vm);
+ char *src_pm_name = sg_host_name(src_pm);
+ char *dst_pm_name = sg_host_name(dst_pm);
+
+ return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
+}
+
+static inline char *get_mig_process_tx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
+{
+ char *vm_name = sg_host_name(vm);
+ char *src_pm_name = sg_host_name(src_pm);
+ char *dst_pm_name = sg_host_name(dst_pm);
+
+ return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
+}
+
+static inline char *get_mig_process_rx_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
+{
+ char *vm_name = sg_host_name(vm);
+ char *src_pm_name = sg_host_name(src_pm);
+ char *dst_pm_name = sg_host_name(dst_pm);
+
+ return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
+}
+
+static inline char *get_mig_task_name(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm, int stage)
+{
+ char *vm_name = sg_host_name(vm);
+ char *src_pm_name = sg_host_name(src_pm);
+ char *dst_pm_name = sg_host_name(dst_pm);
+
+ return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
+}
+
+
+struct migration_session {
+ msg_vm_t vm;
+ msg_host_t src_pm;
+ msg_host_t dst_pm;
+
+ /* The miration_rx process uses mbox_ctl to let the caller of do_migration()
+ * know the completion of the migration. */
+ char *mbox_ctl;
+ /* The migration_rx and migration_tx processes use mbox to transfer migration
+ * data. */
+ char *mbox;
+};
+
+
+static int migration_rx_fun(int argc, char *argv[])
+{
+ XBT_DEBUG("mig: rx_start");
+
+ // The structure has been created in the do_migration function and should only be freed in the same place ;)
+ struct migration_session *ms = MSG_process_get_data(MSG_process_self());
+
+ s_vm_params_t params;
+ simcall_host_get_params(ms->vm, ¶ms);
+
+ int need_exit = 0;
+
+ char *finalize_task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 3);
+
+ int ret = 0;
+ for (;;) {
+ msg_task_t task = NULL;
+ ret = MSG_task_recv(&task, ms->mbox);
+ {
+ if (ret != MSG_OK) {
+ // An error occured, clean the code and return
+ // The owner did not change, hence the task should be only destroyed on the other side
+ xbt_free(finalize_task_name);
+ return 0;
+ }
+ }
+
+ if (strcmp(task->name, finalize_task_name) == 0)
+ need_exit = 1;
+
+ MSG_task_destroy(task);
+
+ if (need_exit)
+ break;
+ }
+
+ // Here Stage 1, 2 and 3 have been performed.
+ // Hence complete the migration
+
+ // Copy the reference to the vm (if SRC crashes now, do_migration will free ms)
+ // This is clearly ugly but I (Adrien) need more time to do something cleaner (actually we should copy the whole ms structure at the begining and free it at the end of each function)
+ msg_vm_t vm = ms->vm;
+ msg_host_t src_pm = ms->src_pm;
+ msg_host_t dst_pm = ms-> dst_pm;
+ msg_host_priv_t priv = sg_host_msg(vm);
+
+// // TODO: we have an issue, if the DST node is turning off during the three next calls, then the VM is in an inconsistent state
+// // I should check with Takahiro in order to make this portion of code atomic
+// /* deinstall the current affinity setting for the CPU */
+// simcall_vm_set_affinity(vm, src_pm, 0);
+//
+// /* Update the vm location */
+// simcall_vm_migrate(vm, dst_pm);
+//
+// /* Resume the VM */
+// simcall_vm_resume(vm);
+//
+ simcall_vm_migratefrom_resumeto(vm, src_pm, dst_pm);
+
+ /* install the affinity setting of the VM on the destination pm */
+ {
+
+ unsigned long affinity_mask = (unsigned long)(uintptr_t) xbt_dict_get_or_null_ext(priv->affinity_mask_db, (char *)dst_pm, sizeof(msg_host_t));
+ simcall_vm_set_affinity(vm, dst_pm, affinity_mask);
+ XBT_DEBUG("set affinity(0x%04lx@%s) for %s", affinity_mask, MSG_host_get_name(dst_pm), MSG_host_get_name(vm));
+ }
+
+ {
+
+ // Now the VM is running on the new host (the migration is completed) (even if the SRC crash)
+ msg_host_priv_t priv = sg_host_msg(vm);
+ priv->is_migrating = 0;
+ XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", ms->vm->key, ms->src_pm->key, ms->dst_pm->key);
+ TRACE_msg_vm_change_host(ms->vm, ms->src_pm, ms->dst_pm);
+ }
+ // Inform the SRC that the migration has been correctly performed
+ {
+ char *task_name = get_mig_task_name(ms->vm, ms->src_pm, ms->dst_pm, 4);
+ msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
+ msg_error_t ret = MSG_task_send(task, ms->mbox_ctl);
+ // xbt_assert(ret == MSG_OK);
+ if(ret == MSG_HOST_FAILURE){
+ // The DST has crashed, this is a problem has the VM since we are not sure whether SRC is considering that the VM has been correctly migrated on the DST node
+ // TODO What does it mean ? What should we do ?
+ MSG_task_destroy(task);
+ } else if(ret == MSG_TRANSFER_FAILURE){
+ // The SRC has crashed, this is not a problem has the VM has been correctly migrated on the DST node
+ MSG_task_destroy(task);
+ }
+
+ xbt_free(task_name);
+ }
+
+
+ xbt_free(finalize_task_name);
+
+ XBT_DEBUG("mig: rx_done");
+
+ return 0;
+}
+
+static void reset_dirty_pages(msg_vm_t vm)
+{
+ msg_host_priv_t priv = sg_host_msg(vm);
+
+ char *key = NULL;
+ xbt_dict_cursor_t cursor = NULL;
+ dirty_page_t dp = NULL;
+ xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
+ double remaining = MSG_task_get_flops_amount(dp->task);
+ dp->prev_clock = MSG_get_clock();
+ dp->prev_remaining = remaining;
+
+ // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
+ }
+}
+
+static void start_dirty_page_tracking(msg_vm_t vm)
+{
+ msg_host_priv_t priv = sg_host_msg(vm);
+ priv->dp_enabled = 1;
+
+ reset_dirty_pages(vm);
+}
+
+static void stop_dirty_page_tracking(msg_vm_t vm)
+{
+ msg_host_priv_t priv = sg_host_msg(vm);
+ priv->dp_enabled = 0;
+}
+
+#if 0
+/* It might be natural that we define dp_rate for each task. But, we will also
+ * have to care about how each task behavior affects the memory update behavior
+ * at the operating system level. It may not be easy to model it with a simple algorithm. */
+double calc_updated_pages(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
+{
+ double computed = dp->prev_remaining - remaining;
+ double duration = clock - dp->prev_clock;
+ double updated = dp->task->dp_rate * computed;
+
+ XBT_INFO("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
+ key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
+ XBT_INFO("%s@%s: updated %f bytes, %f Mbytes/s",
+ key, sg_host_name(vm), updated, updated / duration / 1000 / 1000);
+
+ return updated;
+}
+#endif
+
+static double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
+{
+ double computed = dp->prev_remaining - remaining;
+ double duration = clock - dp->prev_clock;
+
+ XBT_DEBUG("%s@%s: computed %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
+ key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
+
+ return computed;
+}
+
+static double lookup_computed_flop_counts(msg_vm_t vm, int stage_for_fancy_debug, int stage2_round_for_fancy_debug)
+{
+ msg_host_priv_t priv = sg_host_msg(vm);
+ double total = 0;
+
+ char *key = NULL;
+ xbt_dict_cursor_t cursor = NULL;
+ dirty_page_t dp = NULL;
+ xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
+ double remaining = MSG_task_get_flops_amount(dp->task);
+
+ double clock = MSG_get_clock();
+
+ // total += calc_updated_pages(key, vm, dp, remaining, clock);
+ total += get_computed(key, vm, dp, remaining, clock);
+
+ dp->prev_remaining = remaining;
+ dp->prev_clock = clock;
+ }
+
+ total += priv->dp_updated_by_deleted_tasks;
+
+ XBT_DEBUG("mig-stage%d.%d: computed %f flop_counts (including %f by deleted tasks)",
+ stage_for_fancy_debug,
+ stage2_round_for_fancy_debug,
+ total, priv->dp_updated_by_deleted_tasks);
+
+
+
+ priv->dp_updated_by_deleted_tasks = 0;
+
+
+ return total;
+}
+
+// TODO Is this code redundant with the information provided by
+// msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
+void MSG_host_add_task(msg_host_t host, msg_task_t task)
+{
+ msg_host_priv_t priv = sg_host_msg(host);
+ double remaining = MSG_task_get_flops_amount(task);
+ char *key = bprintf("%s-%p", task->name, task);
+
+ dirty_page_t dp = xbt_new0(s_dirty_page, 1);
+ dp->task = task;
+
+ /* It should be okay that we add a task onto a migrating VM. */
+ if (priv->dp_enabled) {
+ dp->prev_clock = MSG_get_clock();
+ dp->prev_remaining = remaining;
+ }
+
+ xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
+ xbt_dict_set(priv->dp_objs, key, dp, NULL);
+ XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_name(host), remaining, priv->dp_enabled);
+
+ xbt_free(key);
+}
+
+void MSG_host_del_task(msg_host_t host, msg_task_t task)
+{
+ msg_host_priv_t priv = sg_host_msg(host);
+
+ char *key = bprintf("%s-%p", task->name, task);
+
+ dirty_page_t dp = xbt_dict_get_or_null(priv->dp_objs, key);
+ xbt_assert(dp->task == task);
+
+ /* If we are in the middle of dirty page tracking, we record how much
+ * computation has been done until now, and keep the information for the
+ * lookup_() function that will called soon. */
+ if (priv->dp_enabled) {
+ double remaining = MSG_task_get_flops_amount(task);
+ double clock = MSG_get_clock();
+ // double updated = calc_updated_pages(key, host, dp, remaining, clock);
+ double updated = get_computed(key, host, dp, remaining, clock);
+
+ priv->dp_updated_by_deleted_tasks += updated;
+ }
+
+ xbt_dict_remove(priv->dp_objs, key);
+ xbt_free(dp);
+
+ XBT_DEBUG("del %s on %s", key, sg_host_name(host));
+
+ xbt_free(key);
+}
+
+
+
+
+static sg_size_t send_migration_data(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm,
+ sg_size_t size, char *mbox, int stage, int stage2_round, double mig_speed, double timeout)
+{
+ sg_size_t sent = 0;
+ char *task_name = get_mig_task_name(vm, src_pm, dst_pm, stage);
+ msg_task_t task = MSG_task_create(task_name, 0, (double)size, NULL);
+
+ /* TODO: clean up */
+
+ double clock_sta = MSG_get_clock();
+
+ msg_error_t ret;
+ if (mig_speed > 0)
+ ret = MSG_task_send_with_timeout_bounded(task, mbox, timeout, mig_speed);
+ else
+ ret = MSG_task_send(task, mbox);
+
+ xbt_free(task_name);
+
+ if (ret == MSG_OK) {
+ sent = size;
+ } else if (ret == MSG_TIMEOUT) {
+ sg_size_t remaining = (sg_size_t)MSG_task_get_remaining_communication(task);
+ sent = size - remaining;
+ XBT_INFO("timeout (%lf s) in sending_migration_data, remaining %llu bytes of %llu",
+ timeout, remaining, size);
+ }
+
+ /* FIXME: why try-and-catch is used here? */
+ if(ret == MSG_HOST_FAILURE){
+ //XBT_INFO("SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
+ MSG_task_destroy(task);
+ THROWF(host_error, 0, "SRC host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
+ }else if(ret == MSG_TRANSFER_FAILURE){
+ //XBT_INFO("DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
+ MSG_task_destroy(task);
+ THROWF(host_error, 0, "DST host failed during migration of %s (stage %d)", sg_host_name(vm), stage);
+ }
+
+ double clock_end = MSG_get_clock();
+ double duration = clock_end - clock_sta;
+ double actual_speed = size / duration;
+
+ if (stage == 2)
+ XBT_DEBUG("mig-stage%d.%d: sent %llu duration %f actual_speed %f (target %f)", stage, stage2_round, size, duration, actual_speed, mig_speed);
+ else
+ XBT_DEBUG("mig-stage%d: sent %llu duration %f actual_speed %f (target %f)", stage, size, duration, actual_speed, mig_speed);
+
+ return sent;