Le vendredi 06 novembre 2026 à 15:38 +0100, Olivier Louvignes a écrit : > Hi Paul, > > On Fri, Nov 06, 2026 at 02:34:08PM +0100, Paul Cercueil wrote: > > > + struct drm_display_mode *mode = &crtc->state->adjusted_mode; > > > > This bothers me a bit, shouldn't it be using 'crtc_state'? > > Yes. v4 takes the mode from crtc_state. > > > > + writel(readl(lcd->base + X1600_LCD_ST), lcd->base + > > > X1600_LCD_CLR_ST); > > > > Not fond of the writel(readl()) nesting here. > > OK, a local in v4, here and in probe. > > > > + writel(val, lcd->base + X1600_LCD_DISP_COM); > > > > This whole block above looks like out of place. Why not do it in > > x1600_lcd_encoder_atomic_mode_set()? > > > > Bonus point, you wouldn't need the 'bus_flags' and 'color_order' variables. > > The controller only takes register writes while its clock runs, and the > clock is off when atomic_mode_set() is called: on a full modeset the > CRTC's atomic_disable() has stopped it, and atomic_enable() starts it > again afterwards. I checked it on the R1: with the clock gated (CLKGR0 > bit 23), every register reads 0x80, and a write to TFT_HSYNC is lost, > the register still holds the old value once the clock is back. > > So moving the block means starting the clock in atomic_mode_set() too, > or leaving it running while the CRTC is off. I'd rather keep the writes > in atomic_enable(), next to clk_prepare_enable(), and look the bus > format up in the new connector state there, so the two fields go > anyway. Would that work for you?
Is there really a point in enabling the clock on-demand? Why not keep it enabled always? It makes the driver more complex and I don't see a real benefit. > > > > + if (pixels > X1600_LCD_MAX_PIXELS) > > > + return MODE_MEM; > > > > The max resolution is 1280x720 so you should check mode->hdisplay / mode- > > > vdisplay individually. Otherwise bogus resolutions like 1920x480 would be > > accepted. > > Right for 1920x480, but 720 per side would refuse the R1's own panel, > which is 480x800 in portrait. The manual only says "frame size up to > 720p". v4 can bound each side by 1280 and keep the 1280x720 pixel count, > would that do? Ok. Then I guess keep the current check. > > > + spin_lock(&lcd->drm.event_lock); > > > > spin_lock() can sleep with CONFIG_PREEMPT_RT so it looks wrong in an IRQ > > handler. I *think* you don't need to lock it here anyway, as the spinlock is > > locked in other places using spin_lock_irq(). > > On PREEMPT_RT the handler doesn't run in hard IRQ context: > force_irqthreads() is always true there, and the IRQ is requested > without IRQF_NO_THREAD, so the handler runs in a thread, where a > spinlock_t is fine (Documentation/locking/locktypes.rst). > > And it needs the lock: drm_crtc_send_vblank_event() wants the caller to > hold event_lock, and lcd->event is shared with queue_event() and > atomic_disable(). Their spin_lock_irq() only keeps the handler off the > CPU they run on. armada, omapdrm and malidp take event_lock the same way > in their handlers. So I'd keep it. Ok. > > You enable it later in the probe, so you could just use > > devm_clk_get_enabled() > > here. > > Probe only enables it to stop what the boot loader left running, and > disables it again before drm_dev_register(). From then on atomic_enable() > and atomic_disable() gate it. devm_clk_get_enabled() would keep it > running for as long as the driver is bound, display off included. Yeah, I forgot that the devm action would cause a clk_disable_unprepare() to be called on unprobe, which wouldn't work here. So ignore that. As for the clock running for as long as the driver is bound, I don't think that's much of a problem. > > I saw a patch from Maxime replacing these - you should be calling > > drm_mode_config_create_initial_state(drm) here (and handle the error). > > Will do, it's in drm-misc-next already. > > I'll leave v3 a few more days for the drm-misc side before sending v4. Great. Cheers, -Paul
