Re: [PATCH v7 2/2] can: m_can: Add hrtimer to generate software interrupt

From: Judith Mendez
Date: Wed May 24 2023 - 17:21:00 EST

Hello Marc,

On 5/23/23 1:35 AM, Marc Kleine-Budde wrote:
On 22.05.2023 21:37:49, Judith Mendez wrote:
Add an hrtimer to MCAN class device. Each MCAN will have its own
hrtimer instantiated if there is no hardware interrupt found in
device tree M_CAN node.

Please add a sentence why you introduce polling mode, i.e. there are
SoCs where the M_CAN interrupt is not available on the CPUs (which are
running Linux).

Sure, I can do that, thanks.


The hrtimer will generate a software interrupt every 1 ms. In
hrtimer callback, we check if there is a transaction pending by
reading a register, then process by calling the isr if there is.

Signed-off-by: Judith Mendez <jm@xxxxxx>
---
Changelog:
v7:
- Clean up m_can_platform.c if/else section after removing poll-interval
- Remove poll-interval from patch description
v6:
- Move hrtimer stop/start function calls to m_can_open and m_can_close to
support power suspend/resume
v5:
- Change dev_dbg to dev_info if hardware interrupt exists and polling
is enabled
v4:
- No changes
v3:
- Create a define for 1 ms polling interval
- Change plarform_get_irq to optional to not print error msg
v2:
- Add functionality to check for 'poll-interval' property in MCAN node
- Add 'polling' flag in driver to check if device is using polling method
- Check for timer polling and hardware interrupt cases, default to
hardware interrupt method
- Change ns_to_ktime() to ms_to_ktime()
---
drivers/net/can/m_can/m_can.c | 33 ++++++++++++++++++++++++--
drivers/net/can/m_can/m_can.h | 4 ++++
drivers/net/can/m_can/m_can_platform.c | 25 ++++++++++++++++---
3 files changed, 57 insertions(+), 5 deletions(-)

diff --git a/drivers/net/can/m_can/m_can.c b/drivers/net/can/m_can/m_can.c
index a5003435802b..f273d989bdff 100644
--- a/drivers/net/can/m_can/m_can.c
+++ b/drivers/net/can/m_can/m_can.c
@@ -11,6 +11,7 @@
#include <linux/bitfield.h>
#include <linux/can/dev.h>
#include <linux/ethtool.h>
+#include <linux/hrtimer.h>
#include <linux/interrupt.h>
#include <linux/io.h>
#include <linux/iopoll.h>
@@ -308,6 +309,9 @@ enum m_can_reg {
#define TX_EVENT_MM_MASK GENMASK(31, 24)
#define TX_EVENT_TXTS_MASK GENMASK(15, 0)
+/* Hrtimer polling interval */
+#define HRTIMER_POLL_INTERVAL 1
+
/* The ID and DLC registers are adjacent in M_CAN FIFO memory,
* and we can save a (potentially slow) bus round trip by combining
* reads and writes to them.
@@ -895,7 +899,7 @@ static int m_can_handle_bus_errors(struct net_device *dev, u32 irqstatus,
netdev_dbg(dev, "Arbitration phase error detected\n");
work_done += m_can_handle_lec_err(dev, lec);
}
-
+

Unrelated change. I've send a separate patch to fix the problem.

Sorry about this, this was not intentional.


if (is_lec_err(dlec)) {
netdev_dbg(dev, "Data phase error detected\n");
work_done += m_can_handle_lec_err(dev, dlec);
@@ -1414,6 +1418,12 @@ static int m_can_start(struct net_device *dev)
m_can_enable_all_interrupts(cdev);
+ if (cdev->polling) {
+ dev_dbg(cdev->dev, "Start hrtimer\n");
+ hrtimer_start(&cdev->hrtimer, ms_to_ktime(HRTIMER_POLL_INTERVAL),
+ HRTIMER_MODE_REL_PINNED);
+ }
+
return 0;
}
@@ -1571,6 +1581,11 @@ static void m_can_stop(struct net_device *dev)
/* disable all interrupts */
m_can_disable_all_interrupts(cdev);
+ if (cdev->polling) {
+ dev_dbg(cdev->dev, "Disabling the hrtimer\n");
+ hrtimer_cancel(&cdev->hrtimer);
+ }
+

This might be a racy. Please move the disabling of the hrtimer before
disabling all interrupts. This makes it also symmetric with respect to
m_can_start().

This makes sense.


/* Set init mode to disengage from the network */
m_can_config_endisable(cdev, true);
@@ -1793,6 +1808,18 @@ static netdev_tx_t m_can_start_xmit(struct sk_buff *skb,
return NETDEV_TX_OK;
}
+static enum hrtimer_restart hrtimer_callback(struct hrtimer *timer)
+{
+ struct m_can_classdev *cdev = container_of(timer, struct
+ m_can_classdev, hrtimer);
+
+ m_can_isr(0, cdev->net);
+
+ hrtimer_forward_now(timer, ms_to_ktime(HRTIMER_POLL_INTERVAL));
+
+ return HRTIMER_RESTART;
+}
+
static int m_can_open(struct net_device *dev)
{
struct m_can_classdev *cdev = netdev_priv(dev);
@@ -1831,9 +1858,11 @@ static int m_can_open(struct net_device *dev)
err = request_threaded_irq(dev->irq, NULL, m_can_isr,
IRQF_ONESHOT,
dev->name, dev);
- } else {
+ } else if (!cdev->polling) {
err = request_irq(dev->irq, m_can_isr, IRQF_SHARED, dev->name,
dev);
+ } else {
+ cdev->hrtimer.function = &hrtimer_callback;

I think you can move this assignment to m_can_class_register(). We only
need to set the function once.

Great idea!


}
if (err < 0) {
diff --git a/drivers/net/can/m_can/m_can.h b/drivers/net/can/m_can/m_can.h
index a839dc71dc9b..e9db5cce4e68 100644
--- a/drivers/net/can/m_can/m_can.h
+++ b/drivers/net/can/m_can/m_can.h
@@ -15,6 +15,7 @@
#include <linux/device.h>
#include <linux/dma-mapping.h>
#include <linux/freezer.h>
+#include <linux/hrtimer.h>
#include <linux/interrupt.h>
#include <linux/io.h>
#include <linux/iopoll.h>
@@ -93,6 +94,9 @@ struct m_can_classdev {
int is_peripheral;
struct mram_cfg mcfg[MRAM_CFG_NUM];
+
+ struct hrtimer hrtimer;
+ bool polling;
};
struct m_can_classdev *m_can_class_allocate_dev(struct device *dev, int sizeof_priv);
diff --git a/drivers/net/can/m_can/m_can_platform.c b/drivers/net/can/m_can/m_can_platform.c
index 94dc82644113..b639c9e645d3 100644
--- a/drivers/net/can/m_can/m_can_platform.c
+++ b/drivers/net/can/m_can/m_can_platform.c
@@ -5,6 +5,7 @@
//
// Copyright (C) 2018-19 Texas Instruments Incorporated - http://www.ti.com/
+#include <linux/hrtimer.h>
#include <linux/phy/phy.h>
#include <linux/platform_device.h>
@@ -96,12 +97,30 @@ static int m_can_plat_probe(struct platform_device *pdev)
goto probe_fail;
addr = devm_platform_ioremap_resource_byname(pdev, "m_can");
- irq = platform_get_irq_byname(pdev, "int0");

I think irq will be uninitialized after this change. Although the
compiler doesn't complain :(

Agreed, I did notice this when testing. But is it an issue?


BTW: I think we don't need the "polling" variable in the priv. We can
make use of "irq". "irq" being 0 means use polling.

True, but is using the polling flag easier for the user to read?


- if (IS_ERR(addr) || irq < 0) {
- ret = -EINVAL;
+ if (IS_ERR(addr)) {
+ ret = PTR_ERR(addr);
goto probe_fail;
}
+ if (device_property_present(mcan_class->dev, "interrupts") ||
+ device_property_present(mcan_class->dev, "interrupt-names")) {
+ irq = platform_get_irq_byname(pdev, "int0");
+ mcan_class->polling = false;
+ if (irq == -EPROBE_DEFER) {
+ ret = -EPROBE_DEFER;
+ goto probe_fail;
+ }
+ if (irq < 0) {
+ ret = -ENXIO;

EINVAL? Wouldn't ENXIO (No such device or address) be more appropriate than
invalid argument?


Please return the original error code.

+ goto probe_fail;
+ }
+ } else {
+ mcan_class->polling = true;
+ dev_dbg(mcan_class->dev, "Polling enabled, initialize hrtimer");
+ hrtimer_init(&mcan_class->hrtimer, CLOCK_MONOTONIC,
+ HRTIMER_MODE_REL_PINNED);
+ }
+
/* message ram could be shared */
res = platform_get_resource_byname(pdev, IORESOURCE_MEM, "message_ram");
if (!res) {

~Judith