| Message ID | 4c9044427b1aab373acd6ac76f0c905e2be79784.1542074855.git.ryder.lee@mediatek.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [resend,1/3] pwm: mediatek: drop flag 'has_clks' | expand |
Hello, On Tue, Nov 13, 2018 at 10:08:22AM +0800, Ryder Lee wrote: > The flag 'has_clks' and related checks are superfluous as the CCF > subsystem does this for you. I'd write instead: Handle optional clocks by using NULL as clk instead of a separate bool field in the device's platform data. There is a semantic difference this patch introduces (i.e. if on mt2712 there are no provided clocks, the driver now successfully binds while before it failed with -ENOENT. And for mt7628 it's the other way round). IMHO this should be noted in the commit log, too. Otherwise it sounds as if this patch was just an optimisation without side effects. > --- > drivers/pwm/pwm-mediatek.c | 20 +++++--------------- > 1 file changed, 5 insertions(+), 15 deletions(-) > > diff --git a/drivers/pwm/pwm-mediatek.c b/drivers/pwm/pwm-mediatek.c > index eb6674c..9400c41 100644 > --- a/drivers/pwm/pwm-mediatek.c > +++ b/drivers/pwm/pwm-mediatek.c > @@ -57,7 +57,6 @@ enum { > struct mtk_pwm_platform_data { > unsigned int num_pwms; > bool pwm45_fixup; > - bool has_clks; > }; > > /** > @@ -87,9 +86,6 @@ static int mtk_pwm_clk_enable(struct pwm_chip *chip, struct pwm_device *pwm) > struct mtk_pwm_chip *pc = to_mtk_pwm_chip(chip); > int ret; > > - if (!pc->soc->has_clks) > - return 0; > - > ret = clk_prepare_enable(pc->clks[MTK_CLK_TOP]); > if (ret < 0) > return ret; > @@ -116,9 +112,6 @@ static void mtk_pwm_clk_disable(struct pwm_chip *chip, struct pwm_device *pwm) > { > struct mtk_pwm_chip *pc = to_mtk_pwm_chip(chip); > > - if (!pc->soc->has_clks) > - return; > - > clk_disable_unprepare(pc->clks[MTK_CLK_PWM1 + pwm->hwpwm]); > clk_disable_unprepare(pc->clks[MTK_CLK_MAIN]); > clk_disable_unprepare(pc->clks[MTK_CLK_TOP]); > @@ -246,12 +239,13 @@ static int mtk_pwm_probe(struct platform_device *pdev) > if (IS_ERR(pc->regs)) > return PTR_ERR(pc->regs); > > - for (i = 0; i < data->num_pwms + 2 && pc->soc->has_clks; i++) { > + for (i = 0; i < data->num_pwms + 2; i++) { > pc->clks[i] = devm_clk_get(&pdev->dev, mtk_pwm_clk_name[i]); > if (IS_ERR(pc->clks[i])) { > - dev_err(&pdev->dev, "clock: %s fail: %ld\n", > - mtk_pwm_clk_name[i], PTR_ERR(pc->clks[i])); > - return PTR_ERR(pc->clks[i]); > + if (PTR_ERR(pc->clks[i]) == -EPROBE_DEFER) > + return -EPROBE_DEFER; > + > + pc->clks[i] = NULL; I'd prefer the following instead: pc->clks[i] = devm_clk_get(&pdev->dev, mtk_pwm_clk_name[i]); if (IS_ERR(pc->clks[i])) { if (PTR_ERR(pc->clks[i]) == -ENODEV) { pc->clks[i] = NULL; } else { if (PTR_ERR(pc->clks[i]) == -EPROBE_DEFER) dev_err(...); return PTR_ERR(pc->clks[i]); } This way you only handle "There is no such clock" and are not ignoring say an IO error. I wonder if it would make sense to introduce functions like: struct clk *clk_get_optional(struct device *dev, const char *id) that return NULL instead of ERR_PTR(-ENODEV). Then the above would simplify to: pc->clks[i] = devm_clk_get_optional(&pdev->dev, mtk_pwm_clk_name[i]); if (IS_ERR(pc->clks[i]) { if (PTR_ERR(pc->clks[i]) == -EPROBE_DEFER) dev_err(...); return PTR_ERR(pc->clks[i]); } (added the clk people to Cc for this question). Best regards Uwe
Quoting Uwe Kleine-König (2018-11-13 01:52:10) > > I wonder if it would make sense to introduce functions like: > > struct clk *clk_get_optional(struct device *dev, const char *id) > > that return NULL instead of ERR_PTR(-ENODEV). > > Then the above would simplify to: > > pc->clks[i] = devm_clk_get_optional(&pdev->dev, mtk_pwm_clk_name[i]); > if (IS_ERR(pc->clks[i]) { > if (PTR_ERR(pc->clks[i]) == -EPROBE_DEFER) > dev_err(...); > return PTR_ERR(pc->clks[i]); > } > > (added the clk people to Cc for this question). > Such a patch is already on the list and not getting much review. http://lkml.kernel.org/r/1535724443-21150-1-git-send-email-phil.edworthy@renesas.com
On Tue, Nov 13, 2018 at 10:08:22AM +0800, Ryder Lee wrote: > The flag 'has_clks' and related checks are superfluous as the CCF > subsystem does this for you. Both of these mechanisms aren't equivalent. While CCF can deal with optional clocks, what the has_clks flag actually means is that the device doesn't need a clock (or doesn't have a clock input) on the devices where it is cleared. So I'd actually be in favor of keeping the has_clks property because it serves as an additional sanity check. For example if you run this driver on an SoC that "has clocks" but if you don't list them in DT, then after this patch the driver will happily continue without clocks, even though it may break completely without those clocks. I've seen SoCs respond to disabled clocks for a hardware block in different ways, in many cases an access to any of the registers will completely hang the CPU. In other cases it may just crash in some other way or give you some sort of machine exception. None of those are good, and make the tiny bit of additional code required to support the has_clks flag very attractive. But that's just my opinion. If you prefer to throw away that safety barrier, be my guest. But if you do, please move this functionality into the clock framework first and then make the driver use it. Thierry
On 14/11/2018 13:47, Thierry Reding wrote: > On Tue, Nov 13, 2018 at 10:08:22AM +0800, Ryder Lee wrote: >> The flag 'has_clks' and related checks are superfluous as the CCF >> subsystem does this for you. > Both of these mechanisms aren't equivalent. While CCF can deal with > optional clocks, what the has_clks flag actually means is that the > device doesn't need a clock (or doesn't have a clock input) on the > devices where it is cleared. > > So I'd actually be in favor of keeping the has_clks property because it > serves as an additional sanity check. For example if you run this driver > on an SoC that "has clocks" but if you don't list them in DT, then after > this patch the driver will happily continue without clocks, even though > it may break completely without those clocks. I've seen SoCs respond to > disabled clocks for a hardware block in different ways, in many cases an > access to any of the registers will completely hang the CPU. In other > cases it may just crash in some other way or give you some sort of > machine exception. None of those are good, and make the tiny bit of > additional code required to support the has_clks flag very attractive. > > But that's just my opinion. If you prefer to throw away that safety > barrier, be my guest. But if you do, please move this functionality into > the clock framework first and then make the driver use it. > > Thierry Hi, sorry for my late response. I added the flag for the legacy MIPS silicon. These SoCs only have a single clock register with a few on/off bits. there is no complex clocktree or scaling. Hence COMMON_CLK is not supported by those SoCs. I fully agree with Thierry, that the flag makes this explicit and the intent was indeed to make sure that on silicon where clocks are required, that they really are listed in OF. This is indeed an extra sanity check and hiding it in an implicit check inside CCF does not feel right. John
On Wed, 2018-11-14 at 14:27 +0100, John Crispin wrote: > On 14/11/2018 13:47, Thierry Reding wrote: > > On Tue, Nov 13, 2018 at 10:08:22AM +0800, Ryder Lee wrote: > >> The flag 'has_clks' and related checks are superfluous as the CCF > >> subsystem does this for you. > > Both of these mechanisms aren't equivalent. While CCF can deal with > > optional clocks, what the has_clks flag actually means is that the > > device doesn't need a clock (or doesn't have a clock input) on the > > devices where it is cleared. > > > > So I'd actually be in favor of keeping the has_clks property because it > > serves as an additional sanity check. For example if you run this driver > > on an SoC that "has clocks" but if you don't list them in DT, then after > > this patch the driver will happily continue without clocks, even though > > it may break completely without those clocks. I've seen SoCs respond to > > disabled clocks for a hardware block in different ways, in many cases an > > access to any of the registers will completely hang the CPU. In other > > cases it may just crash in some other way or give you some sort of > > machine exception. None of those are good, and make the tiny bit of > > additional code required to support the has_clks flag very attractive. > > > > But that's just my opinion. If you prefer to throw away that safety > > barrier, be my guest. But if you do, please move this functionality into > > the clock framework first and then make the driver use it. > > > > Thierry > > Hi, > > sorry for my late response. I added the flag for the legacy MIPS > silicon. These SoCs only have a single clock register with a few on/off > bits. there is no complex clocktree or scaling. Hence COMMON_CLK is not > supported by those SoCs. I fully agree with Thierry, that the flag makes > this explicit and the intent was indeed to make sure that on silicon > where clocks are required, that they really are listed in OF. This is > indeed an extra sanity check and hiding it in an implicit check inside > CCF does not feel right. > > John > Thanks for the detailed information:) I will drop this patch in v1. Ryder
On Wed, Nov 14, 2018 at 01:47:52PM +0100, Thierry Reding wrote: > On Tue, Nov 13, 2018 at 10:08:22AM +0800, Ryder Lee wrote: > > The flag 'has_clks' and related checks are superfluous as the CCF > > subsystem does this for you. > > Both of these mechanisms aren't equivalent. While CCF can deal with > optional clocks, what the has_clks flag actually means is that the > device doesn't need a clock (or doesn't have a clock input) on the > devices where it is cleared. > > So I'd actually be in favor of keeping the has_clks property because it > serves as an additional sanity check. For example if you run this driver > on an SoC that "has clocks" but if you don't list them in DT, then after > this patch the driver will happily continue without clocks, even though > it may break completely without those clocks. I've seen SoCs respond to > disabled clocks for a hardware block in different ways, in many cases an > access to any of the registers will completely hang the CPU. In other > cases it may just crash in some other way or give you some sort of > machine exception. None of those are good, and make the tiny bit of > additional code required to support the has_clks flag very attractive. > > But that's just my opinion. If you prefer to throw away that safety > barrier, be my guest. But if you do, please move this functionality into > the clock framework first and then make the driver use it. The usual policy is: If the things specified in the dt are wrong or incomplete, it's ok to fail however you like. So from a correctness POV I think the change is fine. I don't know about the mips details that John pointed out in a followup to this mail though. Best regards Uwe
On Fri, Nov 16, 2018 at 07:47:48AM +0100, Uwe Kleine-König wrote: > On Wed, Nov 14, 2018 at 01:47:52PM +0100, Thierry Reding wrote: > > On Tue, Nov 13, 2018 at 10:08:22AM +0800, Ryder Lee wrote: > > > The flag 'has_clks' and related checks are superfluous as the CCF > > > subsystem does this for you. > > > > Both of these mechanisms aren't equivalent. While CCF can deal with > > optional clocks, what the has_clks flag actually means is that the > > device doesn't need a clock (or doesn't have a clock input) on the > > devices where it is cleared. > > > > So I'd actually be in favor of keeping the has_clks property because it > > serves as an additional sanity check. For example if you run this driver > > on an SoC that "has clocks" but if you don't list them in DT, then after > > this patch the driver will happily continue without clocks, even though > > it may break completely without those clocks. I've seen SoCs respond to > > disabled clocks for a hardware block in different ways, in many cases an > > access to any of the registers will completely hang the CPU. In other > > cases it may just crash in some other way or give you some sort of > > machine exception. None of those are good, and make the tiny bit of > > additional code required to support the has_clks flag very attractive. > > > > But that's just my opinion. If you prefer to throw away that safety > > barrier, be my guest. But if you do, please move this functionality into > > the clock framework first and then make the driver use it. > > The usual policy is: If the things specified in the dt are > wrong or incomplete, it's ok to fail however you like. So from a > correctness POV I think the change is fine. Erm... that's pretty much what I said. It doesn't necessarily mean that it's the right thing to do, though. If we know that clocks are required and we don't find them in DT, it's better to complain and let the user know exactly what is wrong rather than just let it crash and have them track down the bug without additional information. Thierry
diff --git a/drivers/pwm/pwm-mediatek.c b/drivers/pwm/pwm-mediatek.c index eb6674c..9400c41 100644 --- a/drivers/pwm/pwm-mediatek.c +++ b/drivers/pwm/pwm-mediatek.c @@ -57,7 +57,6 @@ enum { struct mtk_pwm_platform_data { unsigned int num_pwms; bool pwm45_fixup; - bool has_clks; }; /** @@ -87,9 +86,6 @@ static int mtk_pwm_clk_enable(struct pwm_chip *chip, struct pwm_device *pwm) struct mtk_pwm_chip *pc = to_mtk_pwm_chip(chip); int ret; - if (!pc->soc->has_clks) - return 0; - ret = clk_prepare_enable(pc->clks[MTK_CLK_TOP]); if (ret < 0) return ret; @@ -116,9 +112,6 @@ static void mtk_pwm_clk_disable(struct pwm_chip *chip, struct pwm_device *pwm) { struct mtk_pwm_chip *pc = to_mtk_pwm_chip(chip); - if (!pc->soc->has_clks) - return; - clk_disable_unprepare(pc->clks[MTK_CLK_PWM1 + pwm->hwpwm]); clk_disable_unprepare(pc->clks[MTK_CLK_MAIN]); clk_disable_unprepare(pc->clks[MTK_CLK_TOP]); @@ -246,12 +239,13 @@ static int mtk_pwm_probe(struct platform_device *pdev) if (IS_ERR(pc->regs)) return PTR_ERR(pc->regs); - for (i = 0; i < data->num_pwms + 2 && pc->soc->has_clks; i++) { + for (i = 0; i < data->num_pwms + 2; i++) { pc->clks[i] = devm_clk_get(&pdev->dev, mtk_pwm_clk_name[i]); if (IS_ERR(pc->clks[i])) { - dev_err(&pdev->dev, "clock: %s fail: %ld\n", - mtk_pwm_clk_name[i], PTR_ERR(pc->clks[i])); - return PTR_ERR(pc->clks[i]); + if (PTR_ERR(pc->clks[i]) == -EPROBE_DEFER) + return -EPROBE_DEFER; + + pc->clks[i] = NULL; } } @@ -281,25 +275,21 @@ static int mtk_pwm_remove(struct platform_device *pdev) static const struct mtk_pwm_platform_data mt2712_pwm_data = { .num_pwms = 8, .pwm45_fixup = false, - .has_clks = true, }; static const struct mtk_pwm_platform_data mt7622_pwm_data = { .num_pwms = 6, .pwm45_fixup = false, - .has_clks = true, }; static const struct mtk_pwm_platform_data mt7623_pwm_data = { .num_pwms = 5, .pwm45_fixup = true, - .has_clks = true, }; static const struct mtk_pwm_platform_data mt7628_pwm_data = { .num_pwms = 4, .pwm45_fixup = true, - .has_clks = false, }; static const struct of_device_id mtk_pwm_of_match[] = {
The flag 'has_clks' and related checks are superfluous as the CCF subsystem does this for you. Cc: John Crispin <john@phrozen.org> Signed-off-by: Ryder Lee <ryder.lee@mediatek.com> --- drivers/pwm/pwm-mediatek.c | 20 +++++--------------- 1 file changed, 5 insertions(+), 15 deletions(-)