Le 20/06/2022 à 14:48, Tomer Maimon a écrit :
Hi Christophe,
Thanks for your comments.
On Sun, 19 Jun 2022 at 20:14, Christophe JAILLET
<christophe.jaillet-39ZsbGIQGT5GWvitb5QawA@xxxxxxxxxxxxxxxx> wrote:
Le 19/06/2022 à 17:12, Tomer Maimon a écrit :
Nuvoton Arbel BMC NPCM8XX contains an integrated clock controller which
generates and supplies clocks to all modules within the BMC.
Signed-off-by: Tomer Maimon <tmaimon77-Re5JQEeQqe8AvxtiuMwx3w-XMD5yJDbdMQAvxtiuMwx3w@xxxxxxxxxxxxxxxxxxxxxxx>
---
drivers/clk/Kconfig | 6 +
drivers/clk/Makefile | 1 +
drivers/clk/clk-npcm8xx.c | 760 ++++++++++++++++++++++++++++++++++++++
3 files changed, 767 insertions(+)
create mode 100644 drivers/clk/clk-npcm8xx.c
[...]
Hi, below a few comments related to error handling and possible
dev_err_probe() usage to savec a few LoC.
CJ
+static struct clk_hw *
+npcm8xx_clk_register_pll(struct device *dev, void __iomem *pllcon,
+ const char *name, const char *parent_name,
+ unsigned long flags)
+{
+ struct npcm8xx_clk_pll *pll;
+ struct clk_init_data init;
+ struct clk_hw *hw;
+ int ret;
+
+ pll = kzalloc(sizeof(*pll), GFP_KERNEL);
+ if (!pll)
+ return ERR_PTR(-ENOMEM);
+
+ pr_debug("%s reg, name=%s, p=%s\n", __func__, name, parent_name);
+
+ init.name = name;
+ init.ops = &npcm8xx_clk_pll_ops;
+ init.parent_names = &parent_name;
+ init.num_parents = 1;
+ init.flags = flags;
+
+ pll->pllcon = pllcon;
+ pll->hw.init = &init;
+
+ hw = &pll->hw;
+
+ ret = devm_clk_hw_register(dev, hw);
+ if (ret) {
+ kfree(pll);
Hi,
there is no other kfree(() in this patch. It is handled by the framework
once the clk is registered or should there be another kfree() somewhere
or should pll be devm_alloc()'ed?
Could use the devm_kzalloc but the hw is sent to the devm function and
not the pll struct.
My point, as explained below is that some resources are managed (i.e.
allocated by a devm_ function and released by the framework if the probe
fails or if the driver is removed), but not all.
I think that it is not consistent and that maybe the release of the
corresponding resources should be handled in a way or another.
+ return ERR_PTR(ret);
+ }
+
+ return hw;
+}
+
[...]
+static int npcm8xx_clk_probe(struct platform_device *pdev)
+{
+ struct clk_hw_onecell_data *npcm8xx_clk_data;
+ struct device *dev = &pdev->dev;
+ void __iomem *clk_base;
+ struct clk_hw *hw;
+ int i;
+
+ npcm8xx_clk_data = devm_kzalloc(dev, struct_size(npcm8xx_clk_data, hws,
+ NPCM8XX_NUM_CLOCKS),
+ GFP_KERNEL);
+ if (!npcm8xx_clk_data)
+ return -ENOMEM;
+
+ clk_base = devm_platform_ioremap_resource(pdev, 0);
+ if (IS_ERR(clk_base))
+ return PTR_ERR(clk_base);
+
+ npcm8xx_clk_data->num = NPCM8XX_NUM_CLOCKS;
+
+ for (i = 0; i < NPCM8XX_NUM_CLOCKS; i++)
+ npcm8xx_clk_data->hws[i] = ERR_PTR(-EPROBE_DEFER);
+
+ /* Reference 25MHz clock */
+ hw = clk_hw_register_fixed_rate(dev, "refclk", NULL, 0, NPCM8XX_REF_CLK);
Other resoruces are managed, but not this one.
Is it on purpose?
This is the fixed clock - the SoC Reference clock, what do you mean by
not managed?
clk_hw_register_fixed_rate() allocates some memory. If an error occures
later on or if the driver is .remove()'ed, this memory will leak.
(unless I missed something).
Most of your other function calls use some devm_ variant, so the
corresponding resources are released automatically in case of error or
if the driver is unloaed.
Is an error handling path needed or a devm_add_action_or_reset()?
+ if (IS_ERR(hw))
+ return PTR_ERR(hw);
+ npcm8xx_clk_data->hws[NPCM8XX_CLK_REFCLK] = hw;
+
+ /* Register plls */
+ for (i = 0; i < ARRAY_SIZE(npcm8xx_plls); i++) {
+ const struct npcm8xx_clk_pll_data *pll_data = &npcm8xx_plls[i];
+
+ hw = npcm8xx_clk_register_pll(dev, clk_base + pll_data->reg,
+ pll_data->name,
+ pll_data->parent_name,
+ pll_data->flags);
+ if (IS_ERR(hw)) {
+ dev_err(dev, "npcm8xx_clk: Can't register pll\n");
+ return PTR_ERR(hw);
+ }
+
+ if (pll_data->onecell_idx >= 0)
+ npcm8xx_clk_data->hws[pll_data->onecell_idx] = hw;
+ }
+
+ /* Register fixed dividers */
+ hw = devm_clk_hw_register_fixed_factor(dev, NPCM8XX_CLK_S_PLL1_DIV2,
+ NPCM8XX_CLK_S_PLL1, 0, 1, 2);
+ if (IS_ERR(hw)) {
+ dev_err(dev, "npcm8xx_clk: Can't register fixed div\n");
+ return PTR_ERR(hw);
return dev_err_probe()?
it will save online but still enter to a bigger function, I am not
sure that the error will be EPROBE_DEFER, and we have returned the
error in the code.
It is mostly a matter of taste, so if you don't think my proposal
improves anything, just forget about it.
+ }
+
+ hw = devm_clk_hw_register_fixed_factor(dev, NPCM8XX_CLK_S_PLL2_DIV2,
+ NPCM8XX_CLK_S_PLL2, 0, 1, 2);
+ if (IS_ERR(hw)) {
+ dev_err(dev, "npcm8xx_clk: Can't register pll div2\n");
+ return PTR_ERR(hw);
Same here and in other calls below.
+ }
+
+ hw = devm_clk_hw_register_fixed_factor(dev, NPCM8XX_CLK_S_PRE_CLK,
+ NPCM8XX_CLK_S_CPU_MUX, 0, 1, 2);
+ if (IS_ERR(hw)) {
+ dev_err(dev, "npcm8xx_clk: Can't register ckclk div2\n");
+ return PTR_ERR(hw);
+ }
+
+ hw = devm_clk_hw_register_fixed_factor(dev, NPCM8XX_CLK_S_AXI,
+ NPCM8XX_CLK_S_TH, 0, 1, 2);
+ if (IS_ERR(hw)) {
+ dev_err(dev, "npcm8xx_clk: Can't register axi div2\n");
+ return PTR_ERR(hw);
+ }
+
+ hw = devm_clk_hw_register_fixed_factor(dev, NPCM8XX_CLK_S_ATB,
+ NPCM8XX_CLK_S_AXI, 0, 1, 2);
+ if (IS_ERR(hw)) {
+ dev_err(dev, "npcm8xx_clk: Can't register atb div2\n");
+ return PTR_ERR(hw);
+ }
+
+ /* Register muxes */
+ for (i = 0; i < ARRAY_SIZE(npcm8xx_muxes); i++) {
+ const struct npcm8xx_clk_mux_data *mux_data = &npcm8xx_muxes[i];
+
+ hw = clk_hw_register_mux_table(dev, mux_data->name,
+ mux_data->parent_names,
+ mux_data->num_parents,
+ mux_data->flags,
+ clk_base + NPCM8XX_CLKSEL,
+ mux_data->shift,
+ mux_data->mask, 0,
+ mux_data->table,
+ &npcm8xx_clk_lock);
+
Same here. Error handling?
what do you mean?
I mean the same as above. clk_hw_register_mux_table() allocates some
memory. If an error occures later on or if the driver is .remove()'ed,
this memory will leak. (unless I missed something).
Error handling was not a really good wording. You should read "should
the resources allocated by this function be released if an error occures
below"
+ if (IS_ERR(hw)) {
+ dev_err(dev, "npcm8xx_clk: Can't register mux\n");
+ return PTR_ERR(hw);
+ }
+
+ if (mux_data->onecell_idx >= 0)
+ npcm8xx_clk_data->hws[mux_data->onecell_idx] = hw;
+ }
+
+ /* Register clock dividers specified in npcm8xx_divs */
+ for (i = 0; i < ARRAY_SIZE(npcm8xx_divs); i++) {
+ const struct npcm8xx_clk_div_data *div_data = &npcm8xx_divs[i];
+
+ hw = clk_hw_register_divider(dev, div_data->name,
+ div_data->parent_name,
+ div_data->flags,
+ clk_base + div_data->reg,
+ div_data->shift, div_data->width,
+ div_data->clk_divider_flags,
+ &npcm8xx_clk_lock);
devm_clk_hw_register_divider()?
Will do.
Here, the devm_ alternative exist. But my point is the same as above.
(release of resources in cae of error or if the driver is removed)
+ if (IS_ERR(hw)) {
+ dev_err(dev, "npcm8xx_clk: Can't register div table\n");
+ return PTR_ERR(hw);
+ }
+
+ if (div_data->onecell_idx >= 0)
+ npcm8xx_clk_data->hws[div_data->onecell_idx] = hw;
+ }
+
+ return devm_of_clk_add_hw_provider(dev, of_clk_hw_onecell_get,
+ npcm8xx_clk_data);
+}
+
+static const struct of_device_id npcm8xx_clk_dt_ids[] = {
+ { .compatible = "nuvoton,npcm845-clk", },
+ { }
+};
+MODULE_DEVICE_TABLE(of, npcm8xx_clk_dt_ids);
+
+static struct platform_driver npcm8xx_clk_driver = {
+ .probe = npcm8xx_clk_probe,
+ .driver = {
+ .name = "npcm8xx_clk",
+ .of_match_table = npcm8xx_clk_dt_ids,
+ },
+};
+
+static int __init npcm8xx_clk_driver_init(void)
+{
+ return platform_driver_register(&npcm8xx_clk_driver);
+}
+arch_initcall(npcm8xx_clk_driver_init);
+
Best regards,
Tomer
