On Thu, Nov 26, 2020 at 4:07 AM 吳昊澄 Ricky <ricky_wu@xxxxxxxxxxx> wrote: > > > -----Original Message----- > > From: Rafael J. Wysocki [mailto:rafael@xxxxxxxxxx] > > Sent: Wednesday, November 25, 2020 10:04 PM > > To: Bjorn Helgaas; 吳昊澄 Ricky [cut] > > > > +static void rtsx_pci_rtd3_work(struct work_struct *work) > > > > +{ > > > > + struct delayed_work *dwork = to_delayed_work(work); > > > > + struct rtsx_pcr *pcr = container_of(dwork, struct rtsx_pcr, > > rtd3_work); > > > > + > > > > + pcr_dbg(pcr, "--> %s\n", __func__); > > > > + > > > > + while (pcr->pci->dev.power.usage_count.counter > 0) { > > > > + if (pm_runtime_active(&(pcr->pci->dev))) > > > > + pm_runtime_put(&(pcr->pci->dev)); > > > > > > I'm not a runtime PM expert, but this looks fishy. AFAICT this is the > > > only driver in the tree that uses usage_count.counter this way, which > > > is a pretty big hint that this needs a closer look. Cc'd Rafael. > > > > You are right, this is not correct from the PM-runtime POV. > > > > It looks like this attempts to force the PM-runtime usage counter down > > to 0 and it's kind of hard to say why this is done (and it shouldn't > > be done in the first place, because it destroys the usage counter > > balance). > > > > Ricky, is this an attempt to work around an issue of some sort? > > > > Thanks Bjorn and Rafael > I found when we boot up, our dev pcr->pci->dev.power.usage_count.counter always is 2, > Don’t know how to make it to 0 because we need to support D3 and run runtime_suspended callback function > Is there something wrong with us to enable runtime PM? That is possible. If you want it to be enabled by default, you need to call pm_runtime_allow() from the driver at probe time, in addition to pm_runtime_enable(), in the first place, but that only drops one reference, so question is where the other one comes from. Are the pm_runtime_get*() and pm_runtime_put*() calls balanced?
