[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-Id: <20190818182600.3047-6-olteanv@gmail.com>
Date: Sun, 18 Aug 2019 21:26:00 +0300
From: Vladimir Oltean <olteanv@...il.com>
To: broonie@...nel.org, h.feurstein@...il.com, mlichvar@...hat.com,
richardcochran@...il.com, andrew@...n.ch, f.fainelli@...il.com
Cc: linux-spi@...r.kernel.org, netdev@...r.kernel.org,
Vladimir Oltean <olteanv@...il.com>
Subject: [PATCH spi for-5.4 5/5] spi: spi-fsl-dspi: Disable interrupts and preemption during poll mode transfer
While the poll mode helps reduce the overall latency in transmitting a
SPI message in the EOQ and TCFQ modes, the transmission can still have
jitter due to the CPU needing to service interrupts.
The transmission latency does not matter except in situations where the
SPI transfer represents the readout of a POSIX clock. In that case,
even with the byte-level PTP system timestamping in place, a pending IRQ
might find its way to be processed on the local CPU exactly during the
window when the transfer is snapshotted.
Disabling interrupts ensures the above situation never happens. When it
does, it manifests itself as random delay spikes, which throw off the
servo loop of phc2sys and make it lose lock.
Short phc2sys summary after 58 minutes of running without this patch:
offset: min -26251 max 16416 mean -21.8672 std dev 863.416
delay: min 4720 max 57280 mean 5182.49 std dev 1607.19
lost servo lock 3 times
Summary of the same phc2sys service running for 120 minutes with the
patch:
offset: min -378 max 381 mean -0.0083089 std dev 101.495
delay: min 4720 max 5920 mean 5129.38 std dev 154.899
lost servo lock 0 times
Disable interrupts unconditionally if running in poll mode.
Two aspects:
- If the DSPI driver is in IRQ mode, then disabling interrupts becomes a
contradiction in terms. Poll mode is recommendable for predictable
latency.
- In theory it should be possible to disable interrupts only for SPI
transfers that represent an interaction with a POSIX clock. The driver
can sense this by looking at transfer->ptp_sts. However enabling this
unconditionally makes issues much more visible (and not just in fringe
cases), were they to ever appear.
Signed-off-by: Vladimir Oltean <olteanv@...il.com>
---
drivers/spi/spi-fsl-dspi.c | 10 ++++++++++
1 file changed, 10 insertions(+)
diff --git a/drivers/spi/spi-fsl-dspi.c b/drivers/spi/spi-fsl-dspi.c
index ea7169d18e09..c94574a20c8a 100644
--- a/drivers/spi/spi-fsl-dspi.c
+++ b/drivers/spi/spi-fsl-dspi.c
@@ -182,6 +182,8 @@ struct fsl_dspi {
int irq;
struct clk *clk;
+ /* Used to disable IRQs and preemption */
+ spinlock_t lock;
struct ptp_system_timestamp *ptp_sts;
const void *ptp_sts_word_pre;
const void *ptp_sts_word_post;
@@ -739,6 +741,7 @@ static int dspi_transfer_one_message(struct spi_controller *ctlr,
struct spi_device *spi = message->spi;
enum dspi_trans_mode trans_mode;
struct spi_transfer *transfer;
+ unsigned long flags = 0;
int status = 0;
message->actual_length = 0;
@@ -797,6 +800,9 @@ static int dspi_transfer_one_message(struct spi_controller *ctlr,
SPI_FRAME_EBITS(transfer->bits_per_word) |
SPI_CTARE_DTCP(1));
+ if (!dspi->irq)
+ spin_lock_irqsave(&dspi->lock, flags);
+
dspi->take_snapshot_pre = (dspi->tx == dspi->ptp_sts_word_pre);
if (dspi->take_snapshot_pre)
@@ -829,6 +835,9 @@ static int dspi_transfer_one_message(struct spi_controller *ctlr,
do {
status = dspi_poll(dspi);
} while (status == -EINPROGRESS);
+
+ spin_unlock_irqrestore(&dspi->lock, flags);
+
} else if (trans_mode != DSPI_DMA_MODE) {
status = wait_event_interruptible(dspi->waitq,
dspi->waitflags);
@@ -1153,6 +1162,7 @@ static int dspi_probe(struct platform_device *pdev)
}
init_waitqueue_head(&dspi->waitq);
+ spin_lock_init(&dspi->lock);
poll_mode:
if (dspi->devtype_data->trans_mode == DSPI_DMA_MODE) {
--
2.17.1
Powered by blists - more mailing lists