On Tue, Oct 8, 2019 at 2:10 PM Jae Hyun Yoo
<jae.hyun.yoo@xxxxxxxxxxxxxxx> wrote:
>
> Hi Brendan,
>
> On 10/8/2019 1:12 PM, Brendan Higgins wrote:
> > On Mon, Oct 07, 2019 at 04:13:12PM -0700, Jae Hyun Yoo wrote:
> >> Byte mode currently this driver uses makes lots of interrupt call
> >
> > nit: Drop "Byte mode".
>
> 'Byte mode' is one of modes which is described in the datasheet.
>
> Would it be better if I change it like below?
> "This driver uses byte mode that makes lots of interrupt call ..."
Yeah, I think that would probably be clearer.
> >> which isn't good for performance and it makes the driver very
> >> timing sensitive. To improve performance of the driver, this commit
> >> adds buffer mode transfer support which uses I2C SRAM buffer
> >> instead of using a single byte buffer.
> >
> > nit: Please use imperative mood.
>
> I used imperative mood in commit title. The commit message is okay as it
> is.
Hey, that's just what I have been told in the past. I don't actually
feel strongly about it though. If no one else cares, then it is fine.
> >> Signed-off-by: Jae Hyun Yoo <jae.hyun.yoo@xxxxxxxxxxxxxxx>
> >> Tested-by: Tao Ren <taoren@xxxxxx>
> >> ---
> >> drivers/i2c/busses/i2c-aspeed.c | 297 ++++++++++++++++++++++++++++----
> >> 1 file changed, 263 insertions(+), 34 deletions(-)
> >>
> >> diff --git a/drivers/i2c/busses/i2c-aspeed.c b/drivers/i2c/busses/i2c-aspeed.c
> >> index 40f6cf98d32e..37d1a7fa2f87 100644
> >> --- a/drivers/i2c/busses/i2c-aspeed.c
> >> +++ b/drivers/i2c/busses/i2c-aspeed.c
[...]
> >> @@ -238,6 +260,7 @@ static u32 aspeed_i2c_slave_irq(struct aspeed_i2c_bus *bus, u32 irq_status)
> >> {
> >> u32 command, irq_handled = 0;
> >> struct i2c_client *slave = bus->slave;
> >> + int i, len;
> >> u8 value;
> >>
> >> if (!slave)
> >> @@ -260,7 +283,12 @@ static u32 aspeed_i2c_slave_irq(struct aspeed_i2c_bus *bus, u32 irq_status)
> >>
> >> /* Slave was sent something. */
> >> if (irq_status & ASPEED_I2CD_INTR_RX_DONE) {
> >> - value = readl(bus->base + ASPEED_I2C_BYTE_BUF_REG) >> 8;
> >> + if (bus->buf_base &&
> >> + bus->slave_state == ASPEED_I2C_SLAVE_WRITE_RECEIVED &&
> >> + !(irq_status & ASPEED_I2CD_INTR_NORMAL_STOP))
> >
> > I think checking for the buf_base all over the place makes this really
> > complicated and hard to read.
> >
> > It might be better to just split this out and have separate handlers
> > based on what mode the driver is running in.
>
> I think you're saying about splitting this irq handler out to:
> aspeed_i2c_slave_byte_mode_irq()
> aspeed_i2c_slave_buffer_mode_irq()
> aspeed_i2c_slave_dma_mode_irq()
>
> Yes, I can do like that but it will bring us two bad things:
> 1. It makes big chunks of duplicate code because most of interrupt
> handling logic is the same.
> 2. If we are going to change something in irq routine, we need to
> touch all irq routines if the change is commonly used.
>
> I think, the way this patch uses is better.
I think there are other alternatives. For example, I think you could
abstract over the buffer reading mechanism here.
We might have a method on aspeed_i2c_bus called handle_rx_done() or
something like that which could get called here.
I just really don't want to grow the McCabe's complexity of this
function much more, it is really too high as it is. Nevertheless, I am
open to other suggestions on how to improve this function.
> >> + value = readb(bus->buf_base);
> >> + else
> >> + value = readl(bus->base + ASPEED_I2C_BYTE_BUF_REG) >> 8;
> >> /* Handle address frame. */
> >> if (bus->slave_state == ASPEED_I2C_SLAVE_START) {
> >> if (value & 0x1)
> >> @@ -275,6 +303,20 @@ static u32 aspeed_i2c_slave_irq(struct aspeed_i2c_bus *bus, u32 irq_status)
> >>
> >> /* Slave was asked to stop. */
> >> if (irq_status & ASPEED_I2CD_INTR_NORMAL_STOP) {
> >> + if (bus->slave_state == ASPEED_I2C_SLAVE_WRITE_RECEIVED &&
> >> + irq_status & ASPEED_I2CD_INTR_RX_DONE) {
> >> + if (bus->buf_base) {
> >> + len = FIELD_GET(ASPEED_I2CD_BUF_RX_COUNT_MASK,
> >> + readl(bus->base +
> >> + ASPEED_I2C_BUF_CTRL_REG));
> >
> > It looks like you have a lot of improvements in here unrelated to adding
> > support for buffer mode.
> >
> > I really appreciate the improvements, but it makes it harder to
> > understand what buffer features you are adding vs. what
> > improvments/modernizations you are making.
> >
> > Can you split this commit up?
>
> No, this isn't an improvement. This code will not be executed if
> transfer mode is byte mode. This is added because data handling pattern
> is different in buffer mode so the collected data in buffer mode should
> be sent when it recieves RX_DONE.
Oh sorry about that, I saw the switch to the
devm_platform_ioremap_resource below and saw all the FIELD_{GET|PREP}
and assumed that some of them were improvements. If
devm_platform_ioremap_resource is the only one, that's fine.
Actually, would you mind (in a separate commit), update the existing
usages to FIELD_{GET|PREP}? It's kind of jarring going back and forth
between them.
> >> + for (i = 0; i < len; i++) {
> >> + value = readb(bus->buf_base + i);
> >> + i2c_slave_event(slave,
> >> + I2C_SLAVE_WRITE_RECEIVED,
> >> + &value);
> >> + }
> >> + }
> >> + }
> >> irq_handled |= ASPEED_I2CD_INTR_NORMAL_STOP;
> >> bus->slave_state = ASPEED_I2C_SLAVE_STOP;
> >> }
[....]
> >> @@ -990,6 +1180,45 @@ static int aspeed_i2c_probe_bus(struct platform_device *pdev)
> >> bus->get_clk_reg_val = (u32 (*)(struct device *, u32))
> >> match->data;
> >>
> >> + /*
> >> + * Enable I2C SRAM in case of AST2500.
> >> + * SRAM is enabled by default in AST2400 and AST2600.
> >> + */
> >
> > This probe function is already pretty complicated as it is. Can we move
> > this to a helper function (especially since it only applies to the
> > 25xx)?
>
> Okay, that would be better. I'll add this transfer mode setting logic
> as a helper function.
>
> >> + if (of_device_is_compatible(pdev->dev.of_node,
> >> + "aspeed,ast2500-i2c-bus")) {
> >> + struct regmap *gr_regmap = syscon_regmap_lookup_by_compatible("aspeed,ast2500-i2c-gr");
> >
> > So this memory is global, right? It is shared by all the busses?
>
> Yes, this is global register area which can be shared by all busses.
>
> > If I am reading this right, then I think we need to protect so that only
> > one bus is accessing this memory at a time.
>
> It will not be accessed at run time but only at probing time. Since we
> don't use multi-threaded probing, we don't need to protect it.
What if this is loaded as a module?
Also, it seems as though turning on SRAM should only happen once. Is
this correct?
> >> + if (IS_ERR(gr_regmap))
> >> + ret = PTR_ERR(gr_regmap);
> >> + else
> >> + ret = regmap_update_bits(gr_regmap,
> >> + ASPEED_I2CG_GLOBAL_CTRL_REG,
> >> + ASPEED_I2CG_SRAM_BUFFER_EN,
> >> + ASPEED_I2CG_SRAM_BUFFER_EN);
> >> +
> >> + if (ret)
> >> + sram_enabled = false;
> >> + }
> >> +
> >> + if (sram_enabled) {
> >> + struct resource *res = platform_get_resource(pdev,
> >> + IORESOURCE_MEM, 1);
> >> +
> >> + if (res && resource_size(res) >= 2)
> >> + bus->buf_base = devm_ioremap_resource(&pdev->dev, res);
> >> +
> >> + if (!IS_ERR_OR_NULL(bus->buf_base)) {
> >> + bus->buf_size = resource_size(res);
> >> + if (of_device_is_compatible(pdev->dev.of_node,
> >> + "aspeed,ast2400-i2c-bus")) {
> >> + bus->buf_page = ((res->start >> 8) &
> >> + GENMASK(3, 0)) - 8;
> >> + bus->buf_offset = (res->start >> 2) &
> >> + ASPEED_I2CD_BUF_OFFSET_MASK;
> >> + }
> >> + }
> >> + }
[...]
Cheers
