Hi Mason, On Mon, 17 Sep 2018 15:16:18 +0800 masonccyang@xxxxxxxxxxx wrote: > + > +static int mxic_spi_mem_exec_op(struct spi_mem *mem, > + const struct spi_mem_op *op) > +{ > + struct mxic_spi *mxic = spi_master_get_devdata(mem->spi->master); > + int nio = 1, i, ret; > + u32 ss_ctrl; > + u8 addr[8]; > + > + ret = mxic_spi_clk_setup(mem->spi); > + if (ret) > + return ret; Hm, can we change the rate when the clk/PLL is enabled? If not then you'll have to first call pm_runtime_get_sync() to start from a know state, then inside mxic_spi_clk_setup(), call mxic_spi_clk_disable(), change the rate and call mxic_spi_clk_enable(). > + > + ret = mxic_spi_clk_enable(mxic); > + if (ret) > + return ret; Hm, why do you need that one? Isn't it already taken care of by the core when it calls pm_runtime_get_sync() here [1]? > + > + if (mem->spi->mode & (SPI_TX_QUAD | SPI_RX_QUAD)) > + nio = 4; > + else if (mem->spi->mode & (SPI_TX_DUAL | SPI_RX_DUAL)) > + nio = 2; > + > + writel(HC_CFG_NIO(nio) | > + HC_CFG_TYPE(mem->spi->chip_select, HC_CFG_TYPE_SPI_NOR) | > + HC_CFG_SLV_ACT(mem->spi->chip_select) | HC_CFG_IDLE_SIO_LVL(1) | > + HC_CFG_MAN_CS_EN, > + mxic->regs + HC_CFG); > + writel(HC_EN_BIT, mxic->regs + HC_EN); > + > + ss_ctrl = OP_CMD_BYTES(1) | OP_CMD_BUSW(fls(op->cmd.buswidth) - 1); > + > + if (op->addr.nbytes) > + ss_ctrl |= OP_ADDR_BYTES(op->addr.nbytes) | > + OP_ADDR_BUSW(fls(op->addr.buswidth) - 1); > + > + if (op->dummy.nbytes) > + ss_ctrl |= OP_DUMMY_CYC(op->dummy.nbytes); > + > + if (op->data.nbytes) { > + ss_ctrl |= OP_DATA_BUSW(fls(op->data.buswidth) - 1); > + if (op->data.dir == SPI_MEM_DATA_IN) > + ss_ctrl |= OP_READ; > + } > + > + writel(ss_ctrl, mxic->regs + SS_CTRL(mem->spi->chip_select)); > + > + writel(readl(mxic->regs + HC_CFG) | HC_CFG_MAN_CS_ASSERT, > + mxic->regs + HC_CFG); > + > + ret = mxic_spi_data_xfer(mxic, &op->cmd.opcode, NULL, 1); > + if (ret) > + goto out; > + > + for (i = 0; i < op->addr.nbytes; i++) > + addr[i] = op->addr.val >> (8 * (op->addr.nbytes - i - 1)); > + > + ret = mxic_spi_data_xfer(mxic, addr, NULL, op->addr.nbytes); > + if (ret) > + goto out; > + > + ret = mxic_spi_data_xfer(mxic, NULL, NULL, op->dummy.nbytes); > + if (ret) > + goto out; > + > + ret = mxic_spi_data_xfer(mxic, > + op->data.dir == SPI_MEM_DATA_OUT ? > + op->data.buf.out : NULL, > + op->data.dir == SPI_MEM_DATA_IN ? > + op->data.buf.in : NULL, > + op->data.nbytes); > + > +out: > + writel(readl(mxic->regs + HC_CFG) & ~HC_CFG_MAN_CS_ASSERT, > + mxic->regs + HC_CFG); > + writel(0, mxic->regs + HC_EN); > + > + mxic_spi_clk_disable(mxic); > + > + return ret; > +} > + > +static const struct spi_controller_mem_ops mxic_spi_mem_ops = { > + .supports_op = mxic_spi_mem_supports_op, > + .exec_op = mxic_spi_mem_exec_op, > +}; > + > +static void mxic_spi_set_cs(struct spi_device *spi, bool lvl) > +{ > + struct mxic_spi *mxic = spi_master_get_devdata(spi->master); > + > + if (!lvl) { > + if (mxic_spi_clk_setup(spi)) > + return; > + if (mxic_spi_clk_enable(mxic)) > + return; Same here. I'd expect the clks to be enabled by the core [2]. > + writel(readl(mxic->regs + HC_CFG) | HC_CFG_MAN_CS_EN, > + mxic->regs + HC_CFG); > + writel(HC_EN_BIT, mxic->regs + HC_EN); > + writel(readl(mxic->regs + HC_CFG) | HC_CFG_MAN_CS_ASSERT, > + mxic->regs + HC_CFG); > + } else { > + writel(readl(mxic->regs + HC_CFG) & ~HC_CFG_MAN_CS_ASSERT, > + mxic->regs + HC_CFG); > + writel(0, mxic->regs + HC_EN); > + mxic_spi_clk_disable(mxic); > + } > +} > + [...] > + > +static int __maybe_unused mxic_spi_suspend(struct device *dev) > +{ > + struct spi_master *master = dev_get_drvdata(dev); > + struct mxic_spi *mxic = spi_master_get_devdata(master); > + > + clk_disable_unprepare(mxic->ps_clk); Hm, why not putting the ->ps_clk enabling bits in the runtime PM hooks? > + > + spi_master_suspend(master); > + > + return 0; > +} > + [...] > + > +static int __maybe_unused mxic_spi_runtime_suspend(struct device *dev) > +{ > + struct platform_device *pdev = to_platform_device(dev); > + struct spi_master *master = platform_get_drvdata(pdev); > + struct mxic_spi *mxic = spi_master_get_devdata(master); > + > + dev_dbg(dev, "dis-able ps_clock.\n"); ^ disable And I'm not even sure we need this trace here. Looks like some leftovers from your debugging. > + clk_disable_unprepare(mxic->ps_clk); > + > + return 0; > +} > + > +static int __maybe_unused mxic_spi_runtime_resume(struct device *dev) > +{ > + struct platform_device *pdev = to_platform_device(dev); > + struct spi_master *master = platform_get_drvdata(pdev); > + struct mxic_spi *mxic = spi_master_get_devdata(master); > + int ret; > + > + dev_dbg(dev, "en-able ps_clock.\n"); ^enable > + ret = clk_prepare_enable(mxic->ps_clk); > + if (ret) { > + dev_err(dev, "Cannot enable ps_clock.\n"); > + return ret; > + } > + > + return 0; > +} > + > +static const struct dev_pm_ops mxic_spi_dev_pm_ops = { > + SET_RUNTIME_PM_OPS(mxic_spi_runtime_suspend, > + mxic_spi_runtime_resume, NULL) > + SET_SYSTEM_SLEEP_PM_OPS(mxic_spi_suspend, mxic_spi_resume) > +}; > + > +static int mxic_spi_probe(struct platform_device *pdev) > +{ > + struct spi_master *master; > + struct resource *res; > + struct mxic_spi *mxic; > + int ret; > + > + master = spi_alloc_master(&pdev->dev, sizeof(struct mxic_spi)); > + if (!master) > + return -ENOMEM; > + > + platform_set_drvdata(pdev, master); > + > + mxic = spi_master_get_devdata(master); > + > + master->dev.of_node = pdev->dev.of_node; > + > + mxic->ps_clk = devm_clk_get(&pdev->dev, "ps_clk"); > + if (IS_ERR(mxic->ps_clk)) > + return PTR_ERR(mxic->ps_clk); > + > + mxic->send_clk = devm_clk_get(&pdev->dev, "send_clk"); > + if (IS_ERR(mxic->send_clk)) > + return PTR_ERR(mxic->send_clk); > + > + mxic->send_dly_clk = devm_clk_get(&pdev->dev, "send_dly_clk"); > + if (IS_ERR(mxic->send_dly_clk)) > + return PTR_ERR(mxic->send_dly_clk); > + > + res = platform_get_resource_byname(pdev, IORESOURCE_MEM, "regs"); > + mxic->regs = devm_ioremap_resource(&pdev->dev, res); > + if (IS_ERR(mxic->regs)) > + return PTR_ERR(mxic->regs); > + > + ret = clk_prepare_enable(mxic->ps_clk); > + if (ret) > + return ret; > + > + pm_runtime_use_autosuspend(&pdev->dev); > + pm_runtime_set_autosuspend_delay(&pdev->dev, SPI_AUTOSUSPEND_TIMEOUT); > + pm_runtime_set_active(&pdev->dev); > + pm_runtime_enable(&pdev->dev); > + master->auto_runtime_pm = true; > + > + master->num_chipselect = 1; > + master->setup = mxic_spi_setup; > + master->mem_ops = &mxic_spi_mem_ops; > + > + master->set_cs = mxic_spi_set_cs; > + master->transfer_one = mxic_spi_transfer_one; > + master->bits_per_word_mask = SPI_BPW_MASK(8); > + master->mode_bits = SPI_CPOL | SPI_CPHA | > + SPI_RX_DUAL | SPI_TX_DUAL | > + SPI_RX_QUAD | SPI_TX_QUAD; > + > + mxic_spi_hw_init(mxic); > + > + pm_runtime_mark_last_busy(&pdev->dev); > + pm_runtime_put_autosuspend(&pdev->dev); Okay, I'll let others review the runtime PM init/cleanup bits, because I'm not an expert, but I don't remember seeing this pm_runtime_set_active()+pm_runtime_mark_last_busy()+pm_runtime_put_autosuspend() dance in the drivers I worked on. Most probably because they were not using autosuspend. Regards, Boris [1]https://elixir.bootlin.com/linux/v4.19-rc4/source/drivers/spi/spi-mem.c#L210 [2]https://elixir.bootlin.com/linux/v4.19-rc4/source/drivers/spi/spi.c#L1214