On Thu, 13 Jul 2023 10:15:21 +0200, Oswald Buddenhagen wrote: > > On Thu, Jul 13, 2023 at 07:55:31AM +0200, Takashi Iwai wrote: > > On Wed, 12 Jul 2023 16:57:49 +0200, > > Oswald Buddenhagen wrote: > >> > >> The mixer, PCM prepare, MIDI, synth driver, and procfs callbacks are all > >> always invoked with IRQs enabled, so there is no point in saving the > >> state. > >> > >> snd_emu1010_load_firmware_entry() is called from emu1010_firmware_work() > >> and snd_emu10k1_emu1010_init(); the latter from snd_emu10k1_create() and > >> snd_emu10k1_resume(), all of which have IRQs enabled. > >> > >> The voice and memory functions are called from mixed contexts, so they > >> keep the state saving. > >> > >> The low-level functions all keep the state saving, because it's not > >> feasible to keep track of what is called where. > >> > > Wouldn't it make more sense if you replace it with a mutex? > > It'll become more obvious that it's only for non-IRQ context, too. > > > huh? > at least some of the ~six different locks touched by this patch > absolutely _are_ used in irq context. this patch is concerned only > about the specific call sites, where we know that local irqs are > enabled, so we can unconditionally re-enable them rather than > restoring the old state (the latter being a much more expensive > operation). the code already contains precedents for this, and the > complementary optimization of not disabling/restoring irqs where we > know that they are already disabled. > > the reg_lock would be convertible to a mixer_mutex in most mixer > callbacks, but that is an orthogonal question, which is raised in the > next commit. Ah, sorry, I misread as if it were dropping the whole *_irq. Then the patch should be fine. Takashi
