thread_aware now internal & experimental; many fixes; test PoC

Gonna keep the documentation for now

Author: lwaern-intel

Diff for: lib/1.4/dml-builtins.dml

@@ -1128,15 +1128,15 @@ template function_io_memory {
         foreach b in (each function_mapped_bank in (dev)) {
             if (b.function == map_info.function) {
                 local domain_lock_t *lock;
-                #if (dev.thread_aware) SIM_ACQUIRE_CELL(dev.obj, &lock);
+                #if (dev._thread_aware) SIM_ACQUIRE_CELL(dev.obj, &lock);
                 if (!b.use_io_memory) {
                     local map_target_t *mt = SIM_new_map_target(b._bank_obj(),
                                                                  NULL, NULL);
                     if (!mt) {
                         local exception_type_t _exc = SIM_clear_exception();
                         log error: "failed to create map target for %s: %s",
                             SIM_object_name(b._bank_obj()), SIM_last_error();
-                        #if (dev.thread_aware) SIM_RELEASE_CELL(dev.obj, &lock);
+                        #if (dev._thread_aware) SIM_RELEASE_CELL(dev.obj, &lock);
                         return Sim_PE_IO_Not_Taken;
                     }
                     // hack: VT_map_target_access doesn't accept a map_info
@@ -1149,17 +1149,17 @@ template function_io_memory {
                         = VT_map_target_access(mt, mem_op);
                     SIM_set_mem_op_physical_address(mem_op, before);
                     SIM_free_map_target(mt);
-                    #if (dev.thread_aware) SIM_RELEASE_CELL(dev.obj, &lock);
+                    #if (dev._thread_aware) SIM_RELEASE_CELL(dev.obj, &lock);
                     return ret;
                 }
                 if (b.io_memory_access(
                         mem_op,
                         SIM_get_mem_op_physical_address(mem_op) + offset,
                         NULL)) {
-                    #if (dev.thread_aware) SIM_RELEASE_CELL(dev.obj, &lock);
+                    #if (dev._thread_aware) SIM_RELEASE_CELL(dev.obj, &lock);
                     return Sim_PE_No_Exception;
                 } else {
-                    #if (dev.thread_aware) SIM_RELEASE_CELL(dev.obj, &lock);
+                    #if (dev._thread_aware) SIM_RELEASE_CELL(dev.obj, &lock);
                     return Sim_PE_IO_Not_Taken;
                 }
             }
@@ -1380,14 +1380,14 @@ template map_target is (connect, _qname) {
             }
         }
         local domain_lock_t *lock;
-        #if (dev.thread_aware) SIM_ACQUIRE_TARGET(obj, &lock);
+        #if (dev._thread_aware) SIM_ACQUIRE_CELL(obj, &lock);
         local exception_type_t exc = SIM_issue_transaction(map_target, t, addr);
         local bool fail = exc != Sim_PE_No_Exception;
         log info, (fail ? 2 : 4) : "%s%s %d bytes @ 0x%x in %s",
             fail ? "failed to " : "",
             SIM_transaction_is_read(t) ? "read" : fail ? "write" : "wrote",
             SIM_transaction_size(t), addr, SIM_object_name(obj);
-        #if (dev.thread_aware) SIM_RELEASE_TARGET(obj, &lock);
+        #if (dev._thread_aware) SIM_RELEASE_CELL(obj, &lock);
         return exc;
     }
 }
@@ -1451,10 +1451,10 @@ template signal_connect is (connect, post_init) {
 
     method post_init() default {
         local domain_lock_t *lock;
-        #if (dev.thread_aware) SIM_ACQUIRE_CELL(dev.obj, &lock);
+        #if (dev._thread_aware) SIM_ACQUIRE_CELL(dev.obj, &lock);
         if (!SIM_is_restoring_state(dev.obj) && signal.val && signal.high)
             signal.signal_raise();
-        #if (dev.thread_aware) SIM_RELEASE_CELL(dev.obj, &lock);
+        #if (dev._thread_aware) SIM_RELEASE_CELL(dev.obj, &lock);
     }
 
     method set(conf_object_t *obj) default {
@@ -1469,10 +1469,10 @@ template signal_connect is (connect, post_init) {
         //   currently high but the effects of the hotplug has already taken
         //   place so we should NOT treat it as a hotplug.
         local domain_lock_t *lock;
-        #if (dev.thread_aware) SIM_ACQUIRE_CELL(dev.obj, &lock);
+        #if (dev._thread_aware) SIM_ACQUIRE_CELL(dev.obj, &lock);
         local bool hotplug = SIM_object_is_configured(dev.obj)
             && !SIM_is_restoring_state(dev.obj);
-        #if (dev.thread_aware) SIM_RELEASE_CELL(dev.obj, &lock);
+        #if (dev._thread_aware) SIM_RELEASE_CELL(dev.obj, &lock);
         if (hotplug && signal.val && signal.high)
             signal.signal_lower();
         default(obj);

Diff for: lib/1.4/utility.dml

@@ -1269,13 +1269,14 @@ def generate_simple_events_control_methods(device):
     out('{\n', postindent = 1)
     out(crep.structtype(device) + ' *_dev UNUSED = ('
         + crep.structtype(device) + '*)_obj;\n')
-    output_domain_lock_decl('_lock')
-    output_acquire_cell('_obj', '_lock')
-    for key in dml.globals.after_delay_infos:
-        out('SIM_event_cancel_time('
-            + f'SIM_object_clock(_obj), {crep.get_evclass(key)}, _obj, '
-            + '_simple_event_predicate, (lang_void *) &domain);\n')
-    output_release_cell('_obj', '_lock')
+    if dml.globals.after_delay_infos:
+        output_domain_lock_decl('_lock')
+        output_acquire_cell('_obj', '_lock')
+        for key in dml.globals.after_delay_infos:
+            out('SIM_event_cancel_time('
+                + f'SIM_object_clock(_obj), {crep.get_evclass(key)}, _obj, '
+                + '_simple_event_predicate, (lang_void *) &domain);\n')
+        output_release_cell('_obj', '_lock')
 
     site = logging.SimpleSite('<_cancel_simple_events>')
     by_dims = {}
@@ -1336,25 +1337,27 @@ def generate_register_events(device):
     if not events and not dml.globals.after_delay_infos:
         out('return;\n')
     else:
+        flags = 'Sim_EC_No_Serialize' if dml.globals.thread_aware else '0'
         for event in events:
             if (dml.globals.dml_version == (1, 2)
                 and param_str(event, 'timebase') == 'stacked'
                 and event.dimensions > 0):
                 raise ICE(event, "stacked event array not supported")
             for indices in event.all_indices():
-                out('%s%s = SIM_register_event("%s", class, 0, %s);\n'
+                out('%s%s = SIM_register_event("%s", class, %s, %s);\n'
                     % (crep.get_evclass(event),
-                    ''.join('[' + str(i) + ']' for i in indices),
-                    event.logname_anonymized(indices),
-                    ', '.join(
-                        cname
-                        for (_, cname) in event_callbacks(event,
+                       ''.join('[' + str(i) + ']' for i in indices),
+                       event.logname_anonymized(indices),
+                       flags,
+                       ', '.join(
+                           cname
+                           for (_, cname) in event_callbacks(event,
                                                           indices))))
         for (key, info) in dml.globals.after_delay_infos.items():
-            out(('%s = SIM_register_event(%s, class, 0, %s, %s, %s, %s, '
+            out(('%s = SIM_register_event(%s, class, %s, %s, %s, %s, %s, '
                 + 'NULL);\n')
                 % (crep.get_evclass(key), string_literal(info.string_key),
-                   info.cident_callback, '_destroy_simple_event_data',
+                   flags, info.cident_callback, '_destroy_simple_event_data',
                    info.cident_get_value, info.cident_set_value))
     out('}\n\n', preindent = -1)
     splitting_point()

Diff for: py/dml/c_backend.py

@@ -1491,6 +1491,13 @@ def mkobj2(obj, obj_specs, params, each_stmts):
         for (issite, tpl) in obj_spec.templates:
             if tpl.trait:
                 obj_traits.append((issite, tpl.trait))
+            # TODO remove once thread-awareness support is public.
+            # Determining thread-awareness via the _thread_aware param is
+            # cleaner, but doesn't provide is-site info.
+            if (obj.objtype == 'device'
+                and tpl.name == '_thread_aware'
+                and dml.globals.dml_version != (1, 2)):
+                report(WEXPERIMENTAL(issite, 'thread-aware device model'))
 
     for obj_spec in obj_specs:
         for (templates, spec) in obj_spec.in_eachs:
@@ -1956,8 +1963,7 @@ def mkobj2(obj, obj_specs, params, each_stmts):
             mark_method_exported(func, name, export.site)
 
         if dml.globals.dml_version != (1, 2):
-            dml.globals.thread_aware = param_bool_fixup(
-                obj, 'thread_aware', False)
+            dml.globals.thread_aware = param_bool(obj, '_thread_aware')
 
     elif obj.objtype == 'bank':
         set_confidential_object(obj)

Diff for: py/dml/structure.py

@@ -7,12 +7,112 @@ dml 1.4;
 device test;
 
 import "utility.dml";
+import "simics/util/os.dml";
 
 saved int s;
 
-is thread_aware;
+// This test should accomplish two things:
+// 1. Only pass if the device is actually considered thread-aware, and is
+//    entered without the cell domain being taken.
+// 2. Only pass if the device acquires and releases locks correctly
+//
+// Accomplished by two cooperative devices that have a piece of shared state:
+// 1. is accomplished by manipulating the shared state without locking,
+//    relying on sleeps for one thread to observe the changes made by the
+//    other, and performing tests on that state that can only succeed if
+//    the two devices are entered concurrently.
+// 2. is accomplished by manipulations on the shared state while holding
+//    an object lock; then having a device entry to inspect the
+//    state be blocked by an existing entry that modifies it. This blocking is
+//    ensured by also leveraging sleeps. The test is practically guaranteed to
+//    succeed only if object locking is done correctly.
+
+/// WARNING WEXPERIMENTAL
+is _thread_aware;
 
-param init_concurrency_mode = Sim_Concurrency_Mode_Serialized;
+header %{
+    static volatile int32 shared_state = 0;
+%}
+
+extern int32 shared_state;
+
+connect buddy {
+    session int32 *state;
+    param configuration = "optional";
+    interface pulse;
+}
+
+attribute setup_main is write_only_attr {
+    param type = "n";
+    method set(attr_value_t val) throws default {
+        after 1 cycles: event();
+    }
+
+    method event() {
+        os_millisleep(50);
+        // Second resolved (due to the sleep in setup_buddy.event() being
+        // longer)
+
+        // This essentially serves as a check that setup_buddy.event() is
+        // ongoing; and thus the buddy is holding its own lock
+        assert shared_state == 1;
+        // Modify shared state without acquiring lock.
+        shared_state = 2;
+        // Then try to enter the buddy, and become blocked
+        buddy.pulse.pulse();
+    }
+}
+
+attribute setup_buddy is write_only_attr {
+    param type = "n";
+    method set(attr_value_t val) throws default {
+        after 1 cycles: event();
+    }
+
+    method event() {
+        // First resolved (due to the sleep in setup_main.event())
+        shared_state = 1;
+        os_millisleep(100);
+        // Third resolved.
+        assert shared_state == 2;
+        shared_state = 3;
+        // Once this is left, the object lock of the buddy will be released
+    }
+}
+
+implement pulse {
+    method pulse() {
+        // Last resolved, when setup_main.event() becomes unblocked and may
+        // enter the buddy
+        assert shared_state == 3;
+    }
+}
+
+attribute shared_state_attr {
+    param type = "i";
+    method get() -> (attr_value_t) {
+        return SIM_make_attr_int64(shared_state);
+    }
+    method set(attr_value_t val) throws {
+        shared_state = SIM_attr_integer(val);
+    }
+}
+
+bank b;
+port p;
+subdevice sd {
+    subdevice sd {
+        bank b;
+        port p;
+    }
+    bank b;
+    port p;
+}
+
+// Remainder is skeleton code stolen from 1.4/misc/notify_state.
+// We have to do similar tests that object locking is done for the various ways
+// the device may be entered.
+// ... That, or write those tests via ctree_tests.py
 
 header %{
     #include <assert.h>
@@ -127,4 +227,15 @@ implement signal {
 }
 
 
+// port insig is signal_port;
+// connect outsig is signal_connect;
 
+// attribute test_outsig is (pseudo_attr, bool_attr) {
+//     method set(attr_value_t value) throws {
+//         default(value);
+//         if (this.val)
+//             outsig.set_level(1);
+//         else
+//             outsig.set_level(0);
+//     }
+// }