On 8/18/26 5:09 PM, Koichiro Den wrote: Hello Den-san, > I see, CPG_MOD 624. Thanks! Then I am not sure the original writel() is actually > unclocked. It runs before pm_runtime_put(), so runtime PM still keeps that clock > enabled. If "unclocked" means ref clock instead, the new writel() also runs > before clk_bulk_prepare_enable(). Probably I'm missing something again. I would > appreciate it if you could shed light on this. This is clearly my error, the commit message is poorly written. The rcar_gen4_add_dw_pcie_ep() contains two calls which can fail, dw_pcie_ep_init() and dw_pcie_ep_init_registers(). The first one, dw_pcie_ep_init(), internally calls .pre_init callback, which is implemented in rcar_gen4_pcie_ep_pre_init() on R-Car Gen4 . The rcar_gen4_pcie_ep_pre_init() internally calls rcar_gen4_pcie_common_init(), which may succeed or fail. At this point, two possibilities can happen, the fail one is of interest: - If rcar_gen4_pcie_common_init() succeeds , then clk_bulk_prepare_enable() in rcar_gen4_pcie_common_init() did enable all dw->core_clks and those clock are kept enabled on exit from rcar_gen4_pcie_common_init() . The return value from rcar_gen4_pcie_common_init() is 0. OK. - If rcar_gen4_pcie_common_init() failed, then clk_bulk_disable_unprepare() is called in rcar_gen4_pcie_common_init() fail path and dw->core_clks clock are disabled. Return value from rcar_gen4_pcie_common_init() is non-zero. NG. Back in rcar_gen4_pcie_ep_pre_init(): - If rcar_gen4_pcie_ep_pre_init() succeeded , then write PCIEDMAINTSTSEN and enable edma_int , and exit with return value 0, OK. - If rcar_gen4_pcie_common_init() failed, then immediately exit with return value non-zero. Do not write PCIEDMAINTSTSEN and do not enable the edma_interrupts, do depend on previously set content of PCIEDMAINTSTSEN register, which was zeroed out at the beginning of this function instead. NG. Finally, back in rcar_gen4_add_dw_pcie_ep(): - If dw_pcie_ep_init() failed, then dw->core_clks have to be disabled at this point (*), and PCIEDMAINTSTSEN register is 0, therefore, do not call rcar_gen4_pcie_ep_deinit() which writes PCIEDMAINTSTSEN register to 0 (again) and calls rcar_gen4_pcie_common_deinit(), directly call rcar_gen4_pcie_common_deinit() (**). But as I wrote this part, I realized the (*) is not true, and rcar_gen4_pcie_common_deinit() is not the correct function to call in case dw_pcie_ep_init() fails. I think we might need this kind of a patch: " diff --git a/drivers/pci/controller/dwc/pcie-designware-ep.c b/drivers/pci/controller/dwc/pcie-designware-ep.c index de8ee3db43601..a5801a74dec67 100644 --- a/drivers/pci/controller/dwc/pcie-designware-ep.c +++ b/drivers/pci/controller/dwc/pcie-designware-ep.c @@ -1553,7 +1553,7 @@ int dw_pcie_ep_init(struct dw_pcie_ep *ep) ep->page_size); if (ret < 0) { dev_err(dev, "Failed to initialize address space\n"); - return ret; + goto err_deinit; } ep->msi_mem = pci_epc_mem_alloc_addr(epc, &ep->msi_mem_phys, @@ -1568,6 +1568,9 @@ int dw_pcie_ep_init(struct dw_pcie_ep *ep) err_exit_epc_mem: pci_epc_mem_exit(epc); +err_deinit: + if (ep->ops->pre_deinit) + ep->ops->pre_deinit(ep); return ret; } diff --git a/drivers/pci/controller/dwc/pcie-designware.h b/drivers/pci/controller/dwc/pcie-designware.h index 0735ae9409240..733d8b6ad6d2b 100644 --- a/drivers/pci/controller/dwc/pcie-designware.h +++ b/drivers/pci/controller/dwc/pcie-designware.h @@ -475,6 +475,7 @@ struct dw_pcie_rp { struct dw_pcie_ep_ops { int (*pre_init)(struct dw_pcie_ep *ep); + void (*pre_deinit)(struct dw_pcie_ep *ep); int (*init)(struct dw_pcie_ep *ep); int (*raise_irq)(struct dw_pcie_ep *ep, u8 func_no, unsigned int type, u16 interrupt_num); " Then implement .pre_deinit callback for R-Car Gen4 such, that it would assert reset and stop the dw->core_clks by calling rcar_gen4_pcie_common_deinit(). Then we are sure (*) is true. Finally, if dw_pcie_ep_init() call in rcar_gen4_add_dw_pcie_ep() fails, then with the aforementioned deinit implementation, I think it will be possible to simply do the following to cover (**). ret = dw_pcie_ep_init(ep); if (ret) return ret; (I am also attaching the entire example as a diff, compile tested only thus far) What do you think ? > P.S. I also ran two error injection experiments on an S4 Spider. This is also > related to question #2 in my previous comment. > > - When I injected a failure inside common_init(), its error path unwound the > bulk clocks itself. Calling common_deinit() afterwards triggered "already > disabled/unprepared" warnings. I think this ought to be addressed by the pre_deinit above. > - When I injected a failure after pre_init() had succeeded, PCIEDMAINTSTSEN read > back as 0xffff both before and after common_deinit(). I think that with the aforementioned pre_deinit implementation, in case pre_init succeeded and a failure occurred in dw_pcie_ep_init_registers() instead, the PCIEDMAINTSTSEN will be cleared in rcar_gen4_pcie_ep_deinit() correctly. Does that cover this concern or is there another fail path which I missed ? > So I think clean-up needs to know whether pre_init() completed, rather than > always calling common_deinit(). Something like this (field declaration omitted): > > --- a/drivers/pci/controller/dwc/pcie-rcar-gen4.c > +++ b/drivers/pci/controller/dwc/pcie-rcar-gen4.c > @@ -487,17 +487,24 @@ static int rcar_gen4_pcie_ep_pre_init(struct dw_pcie_ep *ep) > struct rcar_gen4_pcie *rcar = to_rcar_gen4_pcie(dw); > int ret; > > ret = rcar_gen4_pcie_common_init(rcar); > if (ret) > return ret; > > writel(PCIEDMAINTSTSEN_INIT, rcar->base + PCIEDMAINTSTSEN); > + rcar->ep_pre_init_done = true; > > return 0; > } > > static void rcar_gen4_pcie_ep_deinit(struct rcar_gen4_pcie *rcar) > { > + if (!rcar->ep_pre_init_done) > + return; > + > + rcar->ep_pre_init_done = false; > writel(0, rcar->base + PCIEDMAINTSTSEN); > rcar_gen4_pcie_common_deinit(rcar); > } > > I may be missing something again, so please just take it with a pinch of salt. Thank you for your feedback, it allowed me to find an error in my train of thoughts, it is very much appreciated. I hope we can also find a good solution. Thank you for your help ! -- Best regards, Marek Vasut