4 * Copyright Red Hat Inc., 2013, 2020
7 * Stefan Hajnoczi <stefanha@redhat.com>
9 * This work is licensed under the terms of the GNU GPL, version 2 or later.
10 * See the COPYING file in the top-level directory.
14 #include "qemu/osdep.h"
15 #include "qom/object.h"
16 #include "qom/object_interfaces.h"
17 #include "qemu/module.h"
18 #include "block/aio.h"
19 #include "block/block.h"
20 #include "sysemu/iothread.h"
21 #include "qapi/error.h"
22 #include "qapi/qapi-commands-misc.h"
23 #include "qemu/error-report.h"
25 #include "qemu/main-loop.h"
27 typedef ObjectClass IOThreadClass
;
29 DECLARE_CLASS_CHECKERS(IOThreadClass
, IOTHREAD
,
33 /* Benchmark results from 2016 on NVMe SSD drives show max polling times around
34 * 16-32 microseconds yield IOPS improvements for both iodepth=1 and iodepth=32
37 #define IOTHREAD_POLL_MAX_NS_DEFAULT 32768ULL
39 #define IOTHREAD_POLL_MAX_NS_DEFAULT 0ULL
42 static void *iothread_run(void *opaque
)
44 IOThread
*iothread
= opaque
;
46 rcu_register_thread();
48 * g_main_context_push_thread_default() must be called before anything
49 * in this new thread uses glib.
51 g_main_context_push_thread_default(iothread
->worker_context
);
52 qemu_set_current_aio_context(iothread
->ctx
);
53 iothread
->thread_id
= qemu_get_thread_id();
54 qemu_sem_post(&iothread
->init_done_sem
);
56 while (iothread
->running
) {
58 * Note: from functional-wise the g_main_loop_run() below can
59 * already cover the aio_poll() events, but we can't run the
60 * main loop unconditionally because explicit aio_poll() here
61 * is faster than g_main_loop_run() when we do not need the
62 * gcontext at all (e.g., pure block layer iothreads). In
63 * other words, when we want to run the gcontext with the
64 * iothread we need to pay some performance for functionality.
66 aio_poll(iothread
->ctx
, true);
69 * We must check the running state again in case it was
70 * changed in previous aio_poll()
72 if (iothread
->running
&& qatomic_read(&iothread
->run_gcontext
)) {
73 g_main_loop_run(iothread
->main_loop
);
77 g_main_context_pop_thread_default(iothread
->worker_context
);
78 rcu_unregister_thread();
82 /* Runs in iothread_run() thread */
83 static void iothread_stop_bh(void *opaque
)
85 IOThread
*iothread
= opaque
;
87 iothread
->running
= false; /* stop iothread_run() */
89 if (iothread
->main_loop
) {
90 g_main_loop_quit(iothread
->main_loop
);
94 void iothread_stop(IOThread
*iothread
)
96 if (!iothread
->ctx
|| iothread
->stopping
) {
99 iothread
->stopping
= true;
100 aio_bh_schedule_oneshot(iothread
->ctx
, iothread_stop_bh
, iothread
);
101 qemu_thread_join(&iothread
->thread
);
104 static void iothread_instance_init(Object
*obj
)
106 IOThread
*iothread
= IOTHREAD(obj
);
108 iothread
->poll_max_ns
= IOTHREAD_POLL_MAX_NS_DEFAULT
;
109 iothread
->thread_id
= -1;
110 qemu_sem_init(&iothread
->init_done_sem
, 0);
111 /* By default, we don't run gcontext */
112 qatomic_set(&iothread
->run_gcontext
, 0);
115 static void iothread_instance_finalize(Object
*obj
)
117 IOThread
*iothread
= IOTHREAD(obj
);
119 iothread_stop(iothread
);
122 * Before glib2 2.33.10, there is a glib2 bug that GSource context
123 * pointer may not be cleared even if the context has already been
124 * destroyed (while it should). Here let's free the AIO context
125 * earlier to bypass that glib bug.
127 * We can remove this comment after the minimum supported glib2
128 * version boosts to 2.33.10. Before that, let's free the
129 * GSources first before destroying any GMainContext.
132 aio_context_unref(iothread
->ctx
);
133 iothread
->ctx
= NULL
;
135 if (iothread
->worker_context
) {
136 g_main_context_unref(iothread
->worker_context
);
137 iothread
->worker_context
= NULL
;
138 g_main_loop_unref(iothread
->main_loop
);
139 iothread
->main_loop
= NULL
;
141 qemu_sem_destroy(&iothread
->init_done_sem
);
144 static void iothread_init_gcontext(IOThread
*iothread
)
148 iothread
->worker_context
= g_main_context_new();
149 source
= aio_get_g_source(iothread_get_aio_context(iothread
));
150 g_source_attach(source
, iothread
->worker_context
);
151 g_source_unref(source
);
152 iothread
->main_loop
= g_main_loop_new(iothread
->worker_context
, TRUE
);
155 static void iothread_complete(UserCreatable
*obj
, Error
**errp
)
157 Error
*local_error
= NULL
;
158 IOThread
*iothread
= IOTHREAD(obj
);
161 iothread
->stopping
= false;
162 iothread
->running
= true;
163 iothread
->ctx
= aio_context_new(errp
);
164 if (!iothread
->ctx
) {
169 * Init one GMainContext for the iothread unconditionally, even if
172 iothread_init_gcontext(iothread
);
174 aio_context_set_poll_params(iothread
->ctx
,
175 iothread
->poll_max_ns
,
177 iothread
->poll_shrink
,
180 error_propagate(errp
, local_error
);
181 aio_context_unref(iothread
->ctx
);
182 iothread
->ctx
= NULL
;
186 /* This assumes we are called from a thread with useful CPU affinity for us
189 thread_name
= g_strdup_printf("IO %s",
190 object_get_canonical_path_component(OBJECT(obj
)));
191 qemu_thread_create(&iothread
->thread
, thread_name
, iothread_run
,
192 iothread
, QEMU_THREAD_JOINABLE
);
195 /* Wait for initialization to complete */
196 while (iothread
->thread_id
== -1) {
197 qemu_sem_wait(&iothread
->init_done_sem
);
203 ptrdiff_t offset
; /* field's byte offset in IOThread struct */
206 static PollParamInfo poll_max_ns_info
= {
207 "poll-max-ns", offsetof(IOThread
, poll_max_ns
),
209 static PollParamInfo poll_grow_info
= {
210 "poll-grow", offsetof(IOThread
, poll_grow
),
212 static PollParamInfo poll_shrink_info
= {
213 "poll-shrink", offsetof(IOThread
, poll_shrink
),
216 static void iothread_get_poll_param(Object
*obj
, Visitor
*v
,
217 const char *name
, void *opaque
, Error
**errp
)
219 IOThread
*iothread
= IOTHREAD(obj
);
220 PollParamInfo
*info
= opaque
;
221 int64_t *field
= (void *)iothread
+ info
->offset
;
223 visit_type_int64(v
, name
, field
, errp
);
226 static void iothread_set_poll_param(Object
*obj
, Visitor
*v
,
227 const char *name
, void *opaque
, Error
**errp
)
229 IOThread
*iothread
= IOTHREAD(obj
);
230 PollParamInfo
*info
= opaque
;
231 int64_t *field
= (void *)iothread
+ info
->offset
;
234 if (!visit_type_int64(v
, name
, &value
, errp
)) {
239 error_setg(errp
, "%s value must be in range [0, %" PRId64
"]",
240 info
->name
, INT64_MAX
);
247 aio_context_set_poll_params(iothread
->ctx
,
248 iothread
->poll_max_ns
,
250 iothread
->poll_shrink
,
255 static void iothread_class_init(ObjectClass
*klass
, void *class_data
)
257 UserCreatableClass
*ucc
= USER_CREATABLE_CLASS(klass
);
258 ucc
->complete
= iothread_complete
;
260 object_class_property_add(klass
, "poll-max-ns", "int",
261 iothread_get_poll_param
,
262 iothread_set_poll_param
,
263 NULL
, &poll_max_ns_info
);
264 object_class_property_add(klass
, "poll-grow", "int",
265 iothread_get_poll_param
,
266 iothread_set_poll_param
,
267 NULL
, &poll_grow_info
);
268 object_class_property_add(klass
, "poll-shrink", "int",
269 iothread_get_poll_param
,
270 iothread_set_poll_param
,
271 NULL
, &poll_shrink_info
);
274 static const TypeInfo iothread_info
= {
275 .name
= TYPE_IOTHREAD
,
276 .parent
= TYPE_OBJECT
,
277 .class_init
= iothread_class_init
,
278 .instance_size
= sizeof(IOThread
),
279 .instance_init
= iothread_instance_init
,
280 .instance_finalize
= iothread_instance_finalize
,
281 .interfaces
= (InterfaceInfo
[]) {
282 {TYPE_USER_CREATABLE
},
287 static void iothread_register_types(void)
289 type_register_static(&iothread_info
);
292 type_init(iothread_register_types
)
294 char *iothread_get_id(IOThread
*iothread
)
296 return g_strdup(object_get_canonical_path_component(OBJECT(iothread
)));
299 AioContext
*iothread_get_aio_context(IOThread
*iothread
)
301 return iothread
->ctx
;
304 static int query_one_iothread(Object
*object
, void *opaque
)
306 IOThreadInfoList
***tail
= opaque
;
310 iothread
= (IOThread
*)object_dynamic_cast(object
, TYPE_IOTHREAD
);
315 info
= g_new0(IOThreadInfo
, 1);
316 info
->id
= iothread_get_id(iothread
);
317 info
->thread_id
= iothread
->thread_id
;
318 info
->poll_max_ns
= iothread
->poll_max_ns
;
319 info
->poll_grow
= iothread
->poll_grow
;
320 info
->poll_shrink
= iothread
->poll_shrink
;
322 QAPI_LIST_APPEND(*tail
, info
);
326 IOThreadInfoList
*qmp_query_iothreads(Error
**errp
)
328 IOThreadInfoList
*head
= NULL
;
329 IOThreadInfoList
**prev
= &head
;
330 Object
*container
= object_get_objects_root();
332 object_child_foreach(container
, query_one_iothread
, &prev
);
336 GMainContext
*iothread_get_g_main_context(IOThread
*iothread
)
338 qatomic_set(&iothread
->run_gcontext
, 1);
339 aio_notify(iothread
->ctx
);
340 return iothread
->worker_context
;
343 IOThread
*iothread_create(const char *id
, Error
**errp
)
347 obj
= object_new_with_props(TYPE_IOTHREAD
,
348 object_get_internal_root(),
351 return IOTHREAD(obj
);
354 void iothread_destroy(IOThread
*iothread
)
356 object_unparent(OBJECT(iothread
));
359 /* Lookup IOThread by its id. Only finds user-created objects, not internal
360 * iothread_create() objects. */
361 IOThread
*iothread_by_id(const char *id
)
363 return IOTHREAD(object_resolve_path_type(id
, TYPE_IOTHREAD
, NULL
));
366 bool qemu_in_iothread(void)
368 return qemu_get_current_aio_context() == qemu_get_aio_context() ?