Re: [PATCH] usb/c67x00: Replace tasklet with work

On Mon, 11 Jan 2021, Hillf Danton wrote:

On Sun, 10 Jan 2021 20:40:50 -0800 Davidlohr Bueso wrote:

quoted

Tasklets have long been deprecated as being too heavy on the system
by running in irq context - and this is not a performance critical
path. If a higher priority process wants to run, it must wait for
the tasklet to finish before doing so.

c67x00_do_work() will now run in process context and have further
concurrency (tasklets being serialized among themselves), but this
is done holding the c67x00->lock, so it should be fine. Furthermore,
this patch fixes the usage of the lock in the callback as otherwise
it would need to be irq-safe.

Can you add a couple of words about the need to be irq-safe because
no lock is taken for scheduling either tasklet or work?

I was refering to the locking in the c67x00_do_work() tasklet callback.
Because it is currently under irq it should be disabling irq (or at least
BH) but after this patch that is no longer the case.

quoted

Signed-off-by: Davidlohr Bueso <redacted>
---
 drivers/usb/c67x00/c67x00-hcd.h   |  2 +-
 drivers/usb/c67x00/c67x00-sched.c | 12 +++++++-----
 2 files changed, 8 insertions(+), 6 deletions(-)

diff --git a/drivers/usb/c67x00/c67x00-hcd.h b/drivers/usb/c67x00/c67x00-hcd.h
index 6b6b04a3fe0f..6332a6b5dce6 100644
--- a/drivers/usb/c67x00/c67x00-hcd.h
+++ b/drivers/usb/c67x00/c67x00-hcd.h

@@ -76,7 +76,7 @@ struct c67x00_hcd {

u16 next_td_addr;
u16 next_buf_addr;

-	struct tasklet_struct tasklet;
+	struct work_struct work;

struct completion endpoint_disable;

diff --git a/drivers/usb/c67x00/c67x00-sched.c b/drivers/usb/c67x00/c67x00-sched.c
index e65f1a0ae80b..af60f4fdd340 100644
--- a/drivers/usb/c67x00/c67x00-sched.c
+++ b/drivers/usb/c67x00/c67x00-sched.c

@@ -1123,24 +1123,26 @@ static void c67x00_do_work(struct c67x00_hcd *c67x00)

 /* -------------------------------------------------------------------------- */

-static void c67x00_sched_tasklet(struct tasklet_struct *t)
+static void c67x00_sched_work(struct work_struct *work)
 {
-	struct c67x00_hcd *c67x00 = from_tasklet(c67x00, t, tasklet);
+	struct c67x00_hcd *c67x00;
+
+	c67x00 = container_of(work, struct c67x00_hcd, work);

c67x00_do_work(c67x00);
 }

 void c67x00_sched_kick(struct c67x00_hcd *c67x00)
 {
-	tasklet_hi_schedule(&c67x00->tasklet);
+        queue_work(system_highpri_wq, &c67x00->work);

Better if one line comment is added for highpri, given this is not a
performance critical path.

I'm not sure the value here, considering the highprio is not being
changed here. There are a few drivers who use highpri workqueue and
care about latencies but they are still not performance critical (to
the overall system that is, which is what I meant by that).

Thanks,
Davidlohr