The thread pool regulates itself: when idle, it kills threads until
empty, when in demand, it creates new threads until full. This behaviour
doesn't play well with latency sensitive workloads where the price of
creating a new thread is too high. For example, when paired with qemu's
'-mlock', or using safety features like SafeStack, creating a new thread
has been measured take multiple milliseconds.
In order to mitigate this let's introduce a new 'EventLoopBase'
property to set the thread pool size. The threads will be created during
the pool's initialization or upon updating the property's value, remain
available during its lifetime regardless of demand, and destroyed upon
freeing it. A properly characterized workload will then be able to
configure the pool to avoid any latency spikes.
Signed-off-by: Nicolas Saenz Julienne <nsaenzju@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Acked-by: Markus Armbruster <armbru@redhat.com>
Message-id: 20220425075723.20019-4-nsaenzju@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Introduce the 'event-loop-base' abstract class, it'll hold the
properties common to all event loops and provide the necessary hooks for
their creation and maintenance. Then have iothread inherit from it.
EventLoopBaseClass is defined as user creatable and provides a hook for
its children to attach themselves to the user creatable class 'complete'
function. It also provides an update_params() callback to propagate
property changes onto its children.
The new 'event-loop-base' class will live in the root directory. It is
built on its own using the 'link_whole' option (there are no direct
function dependencies between the class and its children, it all happens
trough 'constructor' magic). And also imposes new compilation
dependencies:
qom <- event-loop-base <- blockdev (iothread.c)
And in subsequent patches:
qom <- event-loop-base <- qemuutil (util/main-loop.c)
All this forced some amount of reordering in meson.build:
- Moved qom build definition before qemuutil. Doing it the other way
around (i.e. moving qemuutil after qom) isn't possible as a lot of
core libraries that live in between the two depend on it.
- Process the 'hw' subdir earlier, as it introduces files into the
'qom' source set.
No functional changes intended.
Signed-off-by: Nicolas Saenz Julienne <nsaenzju@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Acked-by: Markus Armbruster <armbru@redhat.com>
Message-id: 20220425075723.20019-2-nsaenzju@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Commit 0445409d74 ("iothread: generalize
iothread_set_param/iothread_get_param") moved common code to set and
get IOThread parameters in two new functions.
These functions are called inside callbacks, so we don't need to use an
opaque pointer. Let's replace `void *opaque` parameter with
`IOThreadParamInfo *info`.
Suggested-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
Message-id: 20210727145936.147032-3-sgarzare@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Commit 1793ad0247 ("iothread: add aio-max-batch parameter") added
a new parameter (aio-max-batch) to IOThread and used PollParamInfo
structure to handle it.
Since it is not a parameter of the polling mechanism, we rename the
structure to a more generic IOThreadParamInfo.
Suggested-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
Message-id: 20210727145936.147032-2-sgarzare@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
The `aio-max-batch` parameter will be propagated to AIO engines
and it will be used to control the maximum number of queued requests.
When there are in queue a number of requests equal to `aio-max-batch`,
the engine invokes the system call to forward the requests to the kernel.
This parameter allows us to control the maximum batch size to reduce
the latency that requests might accumulate while queued in the AIO
engine queue.
If `aio-max-batch` is equal to 0 (default value), the AIO engine will
use its default maximum batch size value.
Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Message-id: 20210721094211.69853-3-sgarzare@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Changes in preparation for next patches where we add a new
parameter not related to the poll mechanism.
Let's add two new generic functions (iothread_set_param and
iothread_get_param) that we use to set and get IOThread
parameters.
Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Message-id: 20210721094211.69853-2-sgarzare@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
If we want to wake up a coroutine from a worker thread, aio_co_wake()
currently does not work. In that scenario, aio_co_wake() calls
aio_co_enter(), but there is no current AioContext and therefore
qemu_get_current_aio_context() returns the main thread. aio_co_wake()
then attempts to call aio_context_acquire() instead of going through
aio_co_schedule().
The default case of qemu_get_current_aio_context() was added to cover
synchronous I/O started from the vCPU thread, but the main and vCPU
threads are quite different. The main thread is an I/O thread itself,
only running a more complicated event loop; the vCPU thread instead
is essentially a worker thread that occasionally calls
qemu_mutex_lock_iothread(). It is only in those critical sections
that it acts as if it were the home thread of the main AioContext.
Therefore, this patch detaches qemu_get_current_aio_context() from
iothreads, which is a useless complication. The AioContext pointer
is stored directly in the thread-local variable, including for the
main loop. Worker threads (including vCPU threads) optionally behave
as temporary home threads if they have taken the big QEMU lock,
but if that is not the case they will always schedule coroutines
on remote threads via aio_co_schedule().
With this change, the stub qemu_mutex_iothread_locked() must be changed
from true to false. The previous value of true was needed because the
main thread did not have an AioContext in the thread-local variable,
but now it does have one.
Reported-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20210609122234.544153-1-pbonzini@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Tested-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
[eblake: tweak commit message per Vladimir's review]
Signed-off-by: Eric Blake <eblake@redhat.com>
clang's C11 atomic_fetch_*() functions only take a C11 atomic type
pointer argument. QEMU uses direct types (int, etc) and this causes a
compiler error when a QEMU code calls these functions in a source file
that also included <stdatomic.h> via a system header file:
$ CC=clang CXX=clang++ ./configure ... && make
../util/async.c:79:17: error: address argument to atomic operation must be a pointer to _Atomic type ('unsigned int *' invalid)
Avoid using atomic_*() names in QEMU's atomic.h since that namespace is
used by <stdatomic.h>. Prefix QEMU's APIs with 'q' so that atomic.h
and <stdatomic.h> can co-exist. I checked /usr/include on my machine and
searched GitHub for existing "qatomic_" users but there seem to be none.
This patch was generated using:
$ git grep -h -o '\<atomic\(64\)\?_[a-z0-9_]\+' include/qemu/atomic.h | \
sort -u >/tmp/changed_identifiers
$ for identifier in $(</tmp/changed_identifiers); do
sed -i "s%\<$identifier\>%q$identifier%g" \
$(git grep -I -l "\<$identifier\>")
done
I manually fixed line-wrap issues and misaligned rST tables.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
Acked-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20200923105646.47864-1-stefanha@redhat.com>
object_get_canonical_path_component() returns a malloced copy of a
property name on success, null on failure.
19 of its 25 callers immediately free the returned copy.
Change object_get_canonical_path_component() to return the property
name directly. Since modifying the name would be wrong, adjust the
return type to const char *.
Drop the free from the 19 callers become simpler, add the g_strdup()
to the other six.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Message-Id: <20200714160202.3121879-4-armbru@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
Reviewed-by: Li Qiang <liq3ea@gmail.com>
When all we do with an Error we receive into a local variable is
propagating to somewhere else, we can just as well receive it there
right away. Convert
if (!foo(..., &err)) {
...
error_propagate(errp, err);
...
return ...
}
to
if (!foo(..., errp)) {
...
...
return ...
}
where nothing else needs @err. Coccinelle script:
@rule1 forall@
identifier fun, err, errp, lbl;
expression list args, args2;
binary operator op;
constant c1, c2;
symbol false;
@@
if (
(
- fun(args, &err, args2)
+ fun(args, errp, args2)
|
- !fun(args, &err, args2)
+ !fun(args, errp, args2)
|
- fun(args, &err, args2) op c1
+ fun(args, errp, args2) op c1
)
)
{
... when != err
when != lbl:
when strict
- error_propagate(errp, err);
... when != err
(
return;
|
return c2;
|
return false;
)
}
@rule2 forall@
identifier fun, err, errp, lbl;
expression list args, args2;
expression var;
binary operator op;
constant c1, c2;
symbol false;
@@
- var = fun(args, &err, args2);
+ var = fun(args, errp, args2);
... when != err
if (
(
var
|
!var
|
var op c1
)
)
{
... when != err
when != lbl:
when strict
- error_propagate(errp, err);
... when != err
(
return;
|
return c2;
|
return false;
|
return var;
)
}
@depends on rule1 || rule2@
identifier err;
@@
- Error *err = NULL;
... when != err
Not exactly elegant, I'm afraid.
The "when != lbl:" is necessary to avoid transforming
if (fun(args, &err)) {
goto out
}
...
out:
error_propagate(errp, err);
even though other paths to label out still need the error_propagate().
For an actual example, see sclp_realize().
Without the "when strict", Coccinelle transforms vfio_msix_setup(),
incorrectly. I don't know what exactly "when strict" does, only that
it helps here.
The match of return is narrower than what I want, but I can't figure
out how to express "return where the operand doesn't use @err". For
an example where it's too narrow, see vfio_intx_enable().
Silently fails to convert hw/arm/armsse.c, because Coccinelle gets
confused by ARMSSE being used both as typedef and function-like macro
there. Converted manually.
Line breaks tidied up manually. One nested declaration of @local_err
deleted manually. Preexisting unwanted blank line dropped in
hw/riscv/sifive_e.c.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20200707160613.848843-35-armbru@redhat.com>
Replace
error_setg(&err, ...);
error_propagate(errp, err);
by
error_setg(errp, ...);
Related pattern:
if (...) {
error_setg(&err, ...);
goto out;
}
...
out:
error_propagate(errp, err);
return;
When all paths to label out are that way, replace by
if (...) {
error_setg(errp, ...);
return;
}
and delete the label along with the error_propagate().
When we have at most one other path that actually needs to propagate,
and maybe one at the end that where propagation is unnecessary, e.g.
foo(..., &err);
if (err) {
goto out;
}
...
bar(..., &err);
out:
error_propagate(errp, err);
return;
move the error_propagate() to where it's needed, like
if (...) {
foo(..., &err);
error_propagate(errp, err);
return;
}
...
bar(..., errp);
return;
and transform the error_setg() as above.
In some places, the transformation results in obviously unnecessary
error_propagate(). The next few commits will eliminate them.
Bonus: the elimination of gotos will make later patches in this series
easier to review.
Candidates for conversion tracked down with this Coccinelle script:
@@
identifier err, errp;
expression list args;
@@
- error_setg(&err, args);
+ error_setg(errp, args);
... when != err
error_propagate(errp, err);
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20200707160613.848843-34-armbru@redhat.com>
The previous commit enables conversion of
visit_foo(..., &err);
if (err) {
...
}
to
if (!visit_foo(..., errp)) {
...
}
for visitor functions that now return true / false on success / error.
Coccinelle script:
@@
identifier fun =~ "check_list|input_type_enum|lv_start_struct|lv_type_bool|lv_type_int64|lv_type_str|lv_type_uint64|output_type_enum|parse_type_bool|parse_type_int64|parse_type_null|parse_type_number|parse_type_size|parse_type_str|parse_type_uint64|print_type_bool|print_type_int64|print_type_null|print_type_number|print_type_size|print_type_str|print_type_uint64|qapi_clone_start_alternate|qapi_clone_start_list|qapi_clone_start_struct|qapi_clone_type_bool|qapi_clone_type_int64|qapi_clone_type_null|qapi_clone_type_number|qapi_clone_type_str|qapi_clone_type_uint64|qapi_dealloc_start_list|qapi_dealloc_start_struct|qapi_dealloc_type_anything|qapi_dealloc_type_bool|qapi_dealloc_type_int64|qapi_dealloc_type_null|qapi_dealloc_type_number|qapi_dealloc_type_str|qapi_dealloc_type_uint64|qobject_input_check_list|qobject_input_check_struct|qobject_input_start_alternate|qobject_input_start_list|qobject_input_start_struct|qobject_input_type_any|qobject_input_type_bool|qobject_input_type_bool_keyval|qobject_input_type_int64|qobject_input_type_int64_keyval|qobject_input_type_null|qobject_input_type_number|qobject_input_type_number_keyval|qobject_input_type_size_keyval|qobject_input_type_str|qobject_input_type_str_keyval|qobject_input_type_uint64|qobject_input_type_uint64_keyval|qobject_output_start_list|qobject_output_start_struct|qobject_output_type_any|qobject_output_type_bool|qobject_output_type_int64|qobject_output_type_null|qobject_output_type_number|qobject_output_type_str|qobject_output_type_uint64|start_list|visit_check_list|visit_check_struct|visit_start_alternate|visit_start_list|visit_start_struct|visit_type_.*";
expression list args;
typedef Error;
Error *err;
@@
- fun(args, &err);
- if (err)
+ if (!fun(args, &err))
{
...
}
A few line breaks tidied up manually.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20200707160613.848843-19-armbru@redhat.com>
The only way object_property_add() can fail is when a property with
the same name already exists. Since our property names are all
hardcoded, failure is a programming error, and the appropriate way to
handle it is passing &error_abort.
Same for its variants, except for object_property_add_child(), which
additionally fails when the child already has a parent. Parentage is
also under program control, so this is a programming error, too.
We have a bit over 500 callers. Almost half of them pass
&error_abort, slightly fewer ignore errors, one test case handles
errors, and the remaining few callers pass them to their own callers.
The previous few commits demonstrated once again that ignoring
programming errors is a bad idea.
Of the few ones that pass on errors, several violate the Error API.
The Error ** argument must be NULL, &error_abort, &error_fatal, or a
pointer to a variable containing NULL. Passing an argument of the
latter kind twice without clearing it in between is wrong: if the
first call sets an error, it no longer points to NULL for the second
call. ich9_pm_add_properties(), sparc32_ledma_realize(),
sparc32_dma_realize(), xilinx_axidma_realize(), xilinx_enet_realize()
are wrong that way.
When the one appropriate choice of argument is &error_abort, letting
users pick the argument is a bad idea.
Drop parameter @errp and assert the preconditions instead.
There's one exception to "duplicate property name is a programming
error": the way object_property_add() implements the magic (and
undocumented) "automatic arrayification". Don't drop @errp there.
Instead, rename object_property_add() to object_property_try_add(),
and add the obvious wrapper object_property_add().
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20200505152926.18877-15-armbru@redhat.com>
[Two semantic rebase conflicts resolved]
We were pushing the context until right before running the gmainloop.
Now since we have everything unconditionally, we can move this
earlier.
One benefit is that now it's done even before init_done_sem, so as
long as the iothread user calls iothread_create() and completes, we
know that the thread stack is ready.
Signed-off-by: Peter Xu <peterx@redhat.com>
Message-id: 20190306115532.23025-5-peterx@redhat.com
Message-Id: <20190306115532.23025-5-peterx@redhat.com>
[Tweaked comment wording as discussed with Peter Xu.
--Stefan]
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
In existing code we create the gcontext dynamically at the first
access of the gcontext from caller. That can bring some complexity
and potential races during using iothread. Since the context itself
is not that big a resource, and we won't have millions of iothread,
let's simply create the gcontext unconditionally.
This will also be a preparation work further to move the thread
context push operation earlier than before (now it's only pushed right
before we want to start running the gmainloop).
Removing the g_once since it's not necessary, while introducing a new
run_gcontext boolean to show whether we want to run the gcontext.
Reviewed-by: Marc-André Lureau <marcandre.lureau@redhat.com>
Signed-off-by: Peter Xu <peterx@redhat.com>
Message-id: 20190306115532.23025-3-peterx@redhat.com
Message-Id: <20190306115532.23025-3-peterx@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Lukas reported an hard to reproduce QMP iothread hang on s390 that
QEMU might hang at pthread_join() of the QMP monitor iothread before
quitting:
Thread 1
#0 0x000003ffad10932c in pthread_join
#1 0x0000000109e95750 in qemu_thread_join
at /home/thuth/devel/qemu/util/qemu-thread-posix.c:570
#2 0x0000000109c95a1c in iothread_stop
#3 0x0000000109bb0874 in monitor_cleanup
#4 0x0000000109b55042 in main
While the iothread is still in the main loop:
Thread 4
#0 0x000003ffad0010e4 in ??
#1 0x000003ffad553958 in g_main_context_iterate.isra.19
#2 0x000003ffad553d90 in g_main_loop_run
#3 0x0000000109c9585a in iothread_run
at /home/thuth/devel/qemu/iothread.c:74
#4 0x0000000109e94752 in qemu_thread_start
at /home/thuth/devel/qemu/util/qemu-thread-posix.c:502
#5 0x000003ffad10825a in start_thread
#6 0x000003ffad00dcf2 in thread_start
IMHO it's because there's a race between the main thread and iothread
when stopping the thread in following sequence:
main thread iothread
=========== ==============
aio_poll()
iothread_get_g_main_context
set iothread->worker_context
iothread_stop
schedule iothread_stop_bh
execute iothread_stop_bh [1]
set iothread->running=false
(since main_loop==NULL so
skip to quit main loop.
Note: although main_loop is
NULL but worker_context is
not!)
atomic_read(&iothread->worker_context) [2]
create main_loop object
g_main_loop_run() [3]
pthread_join() [4]
We can see that when execute iothread_stop_bh() at [1] it's possible
that main_loop is still NULL because it's only created until the first
check of the worker_context later at [2]. Then the iothread will hang
in the main loop [3] and it'll starve the main thread too [4].
Here the simple solution should be that we check again the "running"
variable before check against worker_context.
CC: Thomas Huth <thuth@redhat.com>
CC: Dr. David Alan Gilbert <dgilbert@redhat.com>
CC: Stefan Hajnoczi <stefanha@redhat.com>
CC: Lukáš Doktor <ldoktor@redhat.com>
CC: Markus Armbruster <armbru@redhat.com>
CC: Eric Blake <eblake@redhat.com>
CC: Paolo Bonzini <pbonzini@redhat.com>
Reported-by: Lukáš Doktor <ldoktor@redhat.com>
Signed-off-by: Peter Xu <peterx@redhat.com>
Tested-by: Thomas Huth <thuth@redhat.com>
Message-id: 20190129051432.22023-1-peterx@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
-object iothread,id=foo,? will crash qemu:
qemu-system-x86_64:qemu-thread-posix.c:128: qemu_cond_destroy: Assertion `cond->initialized' failed.
Use thread_id != -1 to check if iothread_complete() finished
successfully and the mutex/cond have been initialized.
Signed-off-by: Marc-André Lureau <marcandre.lureau@redhat.com>
Message-Id: <20180821100716.13803-1-marcandre.lureau@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
Free the AIO context earlier than the GMainContext (if we have) to
workaround a glib2 bug that GSource context pointer is not cleared even
if the context has already been destroyed (while it should).
The patch itself only changed the order to destroy the objects, no
functional change at all. Without this workaround, we can encounter
qmp-test hang with oob (and possibly any other use case when iothread is
used with GMainContexts):
#0 0x00007f35ffe45334 in __lll_lock_wait () from /lib64/libpthread.so.0
#1 0x00007f35ffe405d8 in _L_lock_854 () from /lib64/libpthread.so.0
#2 0x00007f35ffe404a7 in pthread_mutex_lock () from /lib64/libpthread.so.0
#3 0x00007f35fc5b9c9d in g_source_unref_internal (source=0x24f0600, context=0x7f35f0000960, have_lock=0) at gmain.c:1685
#4 0x0000000000aa6672 in aio_context_unref (ctx=0x24f0600) at /root/qemu/util/async.c:497
#5 0x000000000065851c in iothread_instance_finalize (obj=0x24f0380) at /root/qemu/iothread.c:129
#6 0x0000000000962d79 in object_deinit (obj=0x24f0380, type=0x242e960) at /root/qemu/qom/object.c:462
#7 0x0000000000962e0d in object_finalize (data=0x24f0380) at /root/qemu/qom/object.c:476
#8 0x0000000000964146 in object_unref (obj=0x24f0380) at /root/qemu/qom/object.c:924
#9 0x0000000000965880 in object_finalize_child_property (obj=0x24ec640, name=0x24efca0 "mon_iothread", opaque=0x24f0380) at /root/qemu/qom/object.c:1436
#10 0x0000000000962c33 in object_property_del_child (obj=0x24ec640, child=0x24f0380, errp=0x0) at /root/qemu/qom/object.c:436
#11 0x0000000000962d26 in object_unparent (obj=0x24f0380) at /root/qemu/qom/object.c:455
#12 0x0000000000658f00 in iothread_destroy (iothread=0x24f0380) at /root/qemu/iothread.c:365
#13 0x00000000004c67a8 in monitor_cleanup () at /root/qemu/monitor.c:4663
#14 0x0000000000669e27 in main (argc=16, argv=0x7ffc8b1ae2f8, envp=0x7ffc8b1ae380) at /root/qemu/vl.c:4749
The glib2 bug is fixed in commit 26056558b ("gmain: allow
g_source_get_context() on destroyed sources", 2012-07-30), so the first
good version is glib2 2.33.10. But we still support building with
glib as old as 2.28, so we need the workaround.
Let's make sure we destroy the GSources first before its owner context
until we drop support for glib older than 2.33.10.
Signed-off-by: Peter Xu <peterx@redhat.com>
Message-Id: <20180409083956.1780-1-peterx@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
