| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* dbus/dbus-marshal.c (_dbus_marshal_validate_arg): fix to avoid
calling _dbus_marshal_validate_arg() for every byte in a byte
array, etc.
* dbus/dbus-message-handler.c: use atomic reference counting to
reduce number of locks slightly; the global lock in here sucks
* dbus/dbus-connection.c
(_dbus_connection_update_dispatch_status_and_unlock): variant of
update_dispatch_status that can be called with lock held; then use
in a couple places to reduce locking/unlocking
(dbus_connection_send): hold the lock over the whole function
instead of acquiring it twice.
* dbus/dbus-timeout.c (_dbus_timeout_new): handle OOM
* bus/connection.c (bus_connections_setup_connection): fix access
to already-freed memory.
* dbus/dbus-connection.c: keep a little cache of linked list
nodes, to avoid using the global linked list alloc lock in the
normal send-message case. Instead we just use the connection lock
that we already have to take.
* dbus/dbus-list.c (_dbus_list_find_last): new function
* dbus/dbus-sysdeps.c (_dbus_atomic_inc, _dbus_atomic_dec):
change to use a struct for the atomic type; fix docs,
they return value before increment, not after increment.
* dbus/dbus-string.c (_dbus_string_append_4_aligned)
(_dbus_string_append_8_aligned): new functions to try to
microoptimize this operation.
(reallocate_for_length): break this out of set_length(), to
improve profile info, and also so we can consider inlining the
set_length() part.
* dbus/dbus-message.c (dbus_message_new_empty_header): init data
strings with some preallocation, cuts down on our calls to realloc
a fair bit. Though if we can get the "move entire string to empty
string" optimization below to kick in here, it would be better.
* dbus/dbus-string.c (_dbus_string_move): just call
_dbus_string_move_len
(_dbus_string_move_len): add a special case for moving
an entire string into an empty string; we can just
swap the string data instead of doing any reallocs.
(_dbus_string_init_preallocated): new function
|
|
|
Write a "test-profile" that does echo client-server with threads;
profile reveals lock contention, memcpy/realloc of buffers, and
UTF-8 validation as hot spots. 20% of lock contention eliminated
with dbus_atomic_inc/dec implementation on x86. Much remaining
contention is global mempool locks for GList and DBusList.
* dbus/dbus-sysdeps.c (_dbus_atomic_inc, _dbus_atomic_dec): add
x86 implementation
* dbus/dbus-connection.c (struct DBusConnection): use
dbus_atomic_t for the reference count
* dbus/dbus-message.c (struct DBusMessage): declare
dbus_atomic_t values as volatile
* configure.in: code to detect ability to use atomic integer
operations in assembly, from GLib patch
* dbus/dbus-internals.c (_dbus_verbose_real): call getpid every
time, tired of it being wrong in threads and forked processes
* glib/test-profile.c: a little program to bounce messages back
and forth between threads and eat CPU
* dbus/dbus-connection.c: add debug spew macros for debugging
thread locks; include config.h at top; fix deadlock in
dbus_connection_flush()
|
