19.06.2019 11:58, Jon Hunter пишет: > > On 18/06/2019 19:26, Dmitry Osipenko wrote: >> 18.06.2019 11:42, Bitan Biswas пишет: >>> tegra_i2c_xfer_msg initiates the I2C transfer in DMA >>> or PIO mode. It involves steps that need FIFO register >>> access, DMA API calls like dma_sync_single_for_device, etc. >>> Tegra I2C ISR has calls to tegra_i2c_empty_rx_fifo in PIO mode >>> and in DMA/PIO mode writes different I2C registers including >>> I2C interrupt status. ISR cannot start processing >>> before the preparation step at tegra_i2c_xfer_msg is complete. >>> Hence, a synchronization between ISR and tegra_i2c_xfer_msg >>> is in place today using spinlock. >> >> Please use full 75 chars per-line, this should make commit message to look better. >> >>> Spinlock busy waits and can add avoidable delays. >>> >>> In this patch needed synchronization is achieved by disabling >>> I2C interrupt during preparation step and enabling interrupt >>> once preparation is over and spinlock is no longer needed. >>> >>> Signed-off-by: Bitan Biswas <bbiswas@xxxxxxxxxx> >>> --- >>> drivers/i2c/busses/i2c-tegra.c | 17 ++++++++--------- >>> 1 file changed, 8 insertions(+), 9 deletions(-) >>> >>> diff --git a/drivers/i2c/busses/i2c-tegra.c b/drivers/i2c/busses/i2c-tegra.c >>> index 6fb545e..ccc7fae 100644 >>> --- a/drivers/i2c/busses/i2c-tegra.c >>> +++ b/drivers/i2c/busses/i2c-tegra.c >>> @@ -240,7 +240,6 @@ struct tegra_i2c_hw_feature { >>> * @bus_clk_rate: current I2C bus clock rate >>> * @clk_divisor_non_hs_mode: clock divider for non-high-speed modes >>> * @is_multimaster_mode: track if I2C controller is in multi-master mode >>> - * @xfer_lock: lock to serialize transfer submission and processing >>> * @tx_dma_chan: DMA transmit channel >>> * @rx_dma_chan: DMA receive channel >>> * @dma_phys: handle to DMA resources >>> @@ -270,8 +269,6 @@ struct tegra_i2c_dev { >>> u32 bus_clk_rate; >>> u16 clk_divisor_non_hs_mode; >>> bool is_multimaster_mode; >>> - /* xfer_lock: lock to serialize transfer submission and processing */ >>> - spinlock_t xfer_lock; >>> struct dma_chan *tx_dma_chan; >>> struct dma_chan *rx_dma_chan; >>> dma_addr_t dma_phys; >>> @@ -835,7 +832,6 @@ static irqreturn_t tegra_i2c_isr(int irq, void *dev_id) >>> >>> status = i2c_readl(i2c_dev, I2C_INT_STATUS); >>> >>> - spin_lock(&i2c_dev->xfer_lock); >>> if (status == 0) { >>> dev_warn(i2c_dev->dev, "irq status 0 %08x %08x %08x\n", >>> i2c_readl(i2c_dev, I2C_PACKET_TRANSFER_STATUS), >>> @@ -935,7 +931,6 @@ static irqreturn_t tegra_i2c_isr(int irq, void *dev_id) >>> >>> complete(&i2c_dev->msg_complete); >>> done: >>> - spin_unlock(&i2c_dev->xfer_lock); >>> return IRQ_HANDLED; >>> } >>> >>> @@ -1054,7 +1049,6 @@ static int tegra_i2c_xfer_msg(struct tegra_i2c_dev *i2c_dev, >>> u32 packet_header; >>> u32 int_mask; >>> unsigned long time_left; >>> - unsigned long flags; >>> size_t xfer_size; >>> u32 *buffer = NULL; >>> int err = 0; >>> @@ -1085,7 +1079,10 @@ static int tegra_i2c_xfer_msg(struct tegra_i2c_dev *i2c_dev, >>> */ >>> xfer_time += DIV_ROUND_CLOSEST(((xfer_size * 9) + 2) * MSEC_PER_SEC, >>> i2c_dev->bus_clk_rate); >>> - spin_lock_irqsave(&i2c_dev->xfer_lock, flags); >>> + if (!i2c_dev->irq_disabled) { >>> + disable_irq_nosync(i2c_dev->irq); >>> + i2c_dev->irq_disabled = true; >>> + } >> >> 1) Peter correctly pointed out in the other email that the disabling should be synced. >> But see more below in 3. >> >> 2) i2c_dev->irq_disabled == true can't ever be the case here because tegra_i2c_init() >> re-enables interrupt in a case of error condition. Hence interrupt always enabled at >> the beginning of the transfer. >> >> 3) In my previous answer I was suggesting to request IRQ in a disabled state, this >> will allow to remove i2c_dev->irq_disabled completely. >> >> Then the tegra_i2c_xfer_msg() will have to enable IRQ after completion of the >> transfer-preparation process and disable IRQ once transfer is done (both success and >> failure cases). This is actually not a bad additional motivation for this patch, to >> keep CPU's interrupt disabled while idling and not to only rely on interrupt masking >> of the I2C hardware. >> >> 4) ISR should simply return IRQ_NONE when interrupt status is 0 and allow kernel core >> to disable the faulty interrupt itself. There will be "unhandled interrupt" error >> message in KMSG log, following the disabling. >> >> 5) In order to request IRQ in a disabled state, the IRQ_NOAUTOEN flag need to be set >> before the requesting, like this: >> >> irq_set_status_flags(irq, IRQ_NOAUTOEN); >> >> devm_request_irq(&pdev->dev, irq...); >> >> In a result of combining 3-5, both i2c_dev->irq_disabled and i2c_dev->irq variables >> become obsolete and could be removed in addition to xfer_lock. That all is a good >> cleanup in my opinion. > > I see, so basically you are simplifying the code by waiting to enable > the interrupt until the transfer is ready and hence you can eliminate > the need for the spinlock. OK, that would make sense. This really needs > to be describe better in the changelog. > > Also what about the tegra_i2c_unmask/mask_irq? Can these be eliminated? Probably, yes. > Finally, what about tegra_i2c_issue_bus_clear()? This requires > interrupts as well. Good points! Bitan, please take this all into account.