Hi, Dan, I agree the best solution is to fix __pm_runtime_resume(). But there are also many cases that assume pm_runtime_get_sync() will change PM usage counter on error. According to my static analysis results, the number of these "right" cases are larger. Adjusting __pm_runtime_resume() directly will introduce more new bugs. Therefore I think we should resolve the "bug" cases individually.
I think that Dmitry's patch is more reasonable than mine. Dinghao "Dan Carpenter" <dan.carpen...@oracle.com>写道： > On Wed, May 20, 2020 at 01:15:44PM +0300, Dmitry Osipenko wrote: > > 20.05.2020 12:51, Dinghao Liu пишет: > > > pm_runtime_get_sync() increments the runtime PM usage counter even > > > it returns an error code. Thus a pairing decrement is needed on > > > the error handling path to keep the counter balanced. > > > > > > Signed-off-by: Dinghao Liu <dinghao....@zju.edu.cn> > > > --- > > > drivers/staging/media/tegra-vde/vde.c | 2 +- > > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > > > diff --git a/drivers/staging/media/tegra-vde/vde.c > > > b/drivers/staging/media/tegra-vde/vde.c > > > index d3e63512a765..dd134a3a15c7 100644 > > > --- a/drivers/staging/media/tegra-vde/vde.c > > > +++ b/drivers/staging/media/tegra-vde/vde.c > > > @@ -777,7 +777,7 @@ static int tegra_vde_ioctl_decode_h264(struct > > > tegra_vde *vde, > > > > > > ret = pm_runtime_get_sync(dev); > > > if (ret < 0) > > > - goto unlock; > > > + goto put_runtime_pm; > > > > > > /* > > > * We rely on the VDE registers reset value, otherwise VDE > > > > > > > Hello Dinghao, > > > > Thank you for the patch. I sent out a similar patch a week ago [1]. > > > > [1] > > https://patchwork.ozlabs.org/project/linux-tegra/patch/20200514210847.9269-2-dig...@gmail.com/ > > > > The pm_runtime_put_noidle() should have the same effect as yours > > variant, although my variant won't change the last_busy RPM time, which > > I think is a bit more appropriate behavior. > > I don't think either patch is correct. The right thing to do is to fix > __pm_runtime_resume() so it doesn't leak a reference count on error. > > The problem is that a lot of functions don't check the return so > possibly we are relying on that behavior. We may need to introduce a > new function which cleans up properly instead of leaking reference > counts? > > Also it's not documented that pm_runtime_get_sync() returns 1 sometimes > on success so it leads to a few bugs. > > drivers/gpu/drm/stm/ltdc.c: ret = pm_runtime_get_sync(ddev->dev); > drivers/gpu/drm/stm/ltdc.c- if (ret) { > -- > drivers/gpu/drm/stm/ltdc.c: ret = pm_runtime_get_sync(ddev->dev); > drivers/gpu/drm/stm/ltdc.c- if (ret) { > > drivers/media/platform/mtk-vcodec/mtk_vcodec_dec_pm.c: ret = > pm_runtime_get_sync(pm->dev); > drivers/media/platform/mtk-vcodec/mtk_vcodec_dec_pm.c- if (ret) > > drivers/media/platform/ti-vpe/cal.c: ret = pm_runtime_get_sync(&pdev->dev); > drivers/media/platform/ti-vpe/cal.c- if (ret) > > drivers/mfd/arizona-core.c: ret = > pm_runtime_get_sync(arizona->dev); > drivers/mfd/arizona-core.c- if (ret != 0) > > drivers/remoteproc/qcom_q6v5_adsp.c: ret = pm_runtime_get_sync(adsp->dev); > drivers/remoteproc/qcom_q6v5_adsp.c- if (ret) > > drivers/spi/spi-img-spfi.c: ret = pm_runtime_get_sync(dev); > drivers/spi/spi-img-spfi.c- if (ret) > > drivers/usb/dwc3/dwc3-pci.c: ret = pm_runtime_get_sync(&dwc3->dev); > drivers/usb/dwc3/dwc3-pci.c- if (ret) > > drivers/watchdog/rti_wdt.c: ret = pm_runtime_get_sync(dev); > drivers/watchdog/rti_wdt.c- if (ret) { > > regards, > dan carpenter > > diff --git a/drivers/base/power/runtime.c b/drivers/base/power/runtime.c > index 99c7da112c95..e280991a977d 100644 > --- a/drivers/base/power/runtime.c > +++ b/drivers/base/power/runtime.c > @@ -1082,6 +1082,9 @@ int __pm_runtime_resume(struct device *dev, int > rpmflags) > retval = rpm_resume(dev, rpmflags); > spin_unlock_irqrestore(&dev->power.lock, flags); > > + if (retval < 0 && rpmflags & RPM_GET_PUT) > + atomic_dec(&dev->power.usage_count); > + > return retval; > } > EXPORT_SYMBOL_GPL(__pm_runtime_resume);
