| Message ID | 20220401120120.7498-1-angelogioacchino.delregno@collabora.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | i2c: mediatek: Optimize master_xfer() and avoid circular locking | expand |
Hi Angelo, On Fri, Apr 01, 2022 at 02:01:20PM +0200, AngeloGioacchino Del Regno wrote: > Especially (but not only) during probe, it may happen that multiple > devices are communicating via i2c (or multiple i2c busses) and > sometimes while others are probing asynchronously. > For example, a Cr50 TPM may be filling entropy (or userspace may be > reading random data) while the rt5682 (i2c) codec driver reads/sets > some registers, like while getting/setting a clock's rate, which > happens both during probe and during system operation. > > In this driver, the mtk_i2c_transfer() function (which is the i2c > .master_xfer() callback) was granularly managing the clocks by > performing a clk_bulk_prepare_enable() to start them and its inverse. > This is not only creating possible circular locking dependencies in > the some cases (like former explaination), but it's also suboptimal, s/explaination/explanation/ > as clk_core prepare/unprepare operations are using mutex locking, > which creates a bit of unwanted overhead (for example, i2c trackpads > will call master_xfer() every few milliseconds!). > > With this commit, we avoid both the circular locking and additional > overhead by changing how we handle the clocks in this driver: > - Prepare the clocks during probe (and PM resume) > - Enable/disable clocks in mtk_i2c_transfer() > - Unprepare the clocks only for driver removal (and PM suspend) > > For the sake of providing a full explaination: during probe, the Ditto. > clocks are not only prepared but also enabled, as this is needed > for somehardware initialization but, after that, we are disabling s/somehardware/some hardware/ > but not unpreparing them, leaving an expected state for the > aforementioned clock handling strategy. > > Signed-off-by: AngeloGioacchino Del Regno <angelogioacchino.delregno@collabora.com> I couldn't see any performance improvements on mt8192-asurada-spherion with this, although I didn't do any serious benchmark testing. Besides, on spherion the TPM is instead wired through SPI, so I guess most of the boot overhead wouldn't be visible there. In any case it did get rid of the frequent deadlocks when booting on spherion, which is already great on its own, so Tested-by: Nícolas F. R. A. Prado <nfraprado@collabora.com> Thanks, Nícolas
diff --git a/drivers/i2c/busses/i2c-mt65xx.c b/drivers/i2c/busses/i2c-mt65xx.c index f651d3e124d6..bdecb78bfc26 100644 --- a/drivers/i2c/busses/i2c-mt65xx.c +++ b/drivers/i2c/busses/i2c-mt65xx.c @@ -1177,7 +1177,7 @@ static int mtk_i2c_transfer(struct i2c_adapter *adap, int left_num = num; struct mtk_i2c *i2c = i2c_get_adapdata(adap); - ret = clk_bulk_prepare_enable(I2C_MT65XX_CLK_MAX, i2c->clocks); + ret = clk_bulk_enable(I2C_MT65XX_CLK_MAX, i2c->clocks); if (ret) return ret; @@ -1231,7 +1231,7 @@ static int mtk_i2c_transfer(struct i2c_adapter *adap, ret = num; err_exit: - clk_bulk_disable_unprepare(I2C_MT65XX_CLK_MAX, i2c->clocks); + clk_bulk_disable(I2C_MT65XX_CLK_MAX, i2c->clocks); return ret; } @@ -1412,7 +1412,7 @@ static int mtk_i2c_probe(struct platform_device *pdev) return ret; } mtk_i2c_init_hw(i2c); - clk_bulk_disable_unprepare(I2C_MT65XX_CLK_MAX, i2c->clocks); + clk_bulk_disable(I2C_MT65XX_CLK_MAX, i2c->clocks); ret = devm_request_irq(&pdev->dev, irq, mtk_i2c_irq, IRQF_NO_SUSPEND | IRQF_TRIGGER_NONE, @@ -1439,6 +1439,8 @@ static int mtk_i2c_remove(struct platform_device *pdev) i2c_del_adapter(&i2c->adap); + clk_bulk_unprepare(I2C_MT65XX_CLK_MAX, i2c->clocks); + return 0; } @@ -1448,6 +1450,7 @@ static int mtk_i2c_suspend_noirq(struct device *dev) struct mtk_i2c *i2c = dev_get_drvdata(dev); i2c_mark_adapter_suspended(&i2c->adap); + clk_bulk_unprepare(I2C_MT65XX_CLK_MAX, i2c->clocks); return 0; } @@ -1465,7 +1468,7 @@ static int mtk_i2c_resume_noirq(struct device *dev) mtk_i2c_init_hw(i2c); - clk_bulk_disable_unprepare(I2C_MT65XX_CLK_MAX, i2c->clocks); + clk_bulk_disable(I2C_MT65XX_CLK_MAX, i2c->clocks); i2c_mark_adapter_resumed(&i2c->adap);
Especially (but not only) during probe, it may happen that multiple devices are communicating via i2c (or multiple i2c busses) and sometimes while others are probing asynchronously. For example, a Cr50 TPM may be filling entropy (or userspace may be reading random data) while the rt5682 (i2c) codec driver reads/sets some registers, like while getting/setting a clock's rate, which happens both during probe and during system operation. In this driver, the mtk_i2c_transfer() function (which is the i2c .master_xfer() callback) was granularly managing the clocks by performing a clk_bulk_prepare_enable() to start them and its inverse. This is not only creating possible circular locking dependencies in the some cases (like former explaination), but it's also suboptimal, as clk_core prepare/unprepare operations are using mutex locking, which creates a bit of unwanted overhead (for example, i2c trackpads will call master_xfer() every few milliseconds!). With this commit, we avoid both the circular locking and additional overhead by changing how we handle the clocks in this driver: - Prepare the clocks during probe (and PM resume) - Enable/disable clocks in mtk_i2c_transfer() - Unprepare the clocks only for driver removal (and PM suspend) For the sake of providing a full explaination: during probe, the clocks are not only prepared but also enabled, as this is needed for somehardware initialization but, after that, we are disabling but not unpreparing them, leaving an expected state for the aforementioned clock handling strategy. Signed-off-by: AngeloGioacchino Del Regno <angelogioacchino.delregno@collabora.com> --- This patch was successfully tested on various chromebooks with different MediaTek SoCs. All chromebooks are using a Cr50 TPM, trackpad and/or touchscreen over i2c. Some of them are using a i2c audio codec as well. drivers/i2c/busses/i2c-mt65xx.c | 11 +++++++---- 1 file changed, 7 insertions(+), 4 deletions(-)