d/freebsd-dev
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

In am335x_dmtpps, use a spin mutex to interlock between PPS capture and PPS

Browse Source 
ioctl(2) handling.  This allows doing the pps_event() work in the polling
routine, instead of using a taskqueue task to do that work.

Also, add PNPINFO, and switch to using make_dev_s() to create the cdev.

Using a spin mutex and calling pps_event() from the polling function works
around the situation which requires more than 2 sets of timecounter
timehands in a single-core system to get reliable PPS capture.  That problem
would happen when a single-core system is idle in cpu_idle() then gets woken
up with an event timer event which was scheduled to handle a hardclock tick.
That processing path would end up calling tc_windup 3 or 4 times between
when the tc polling function was called and when the taskqueue task would
eventually run, and with only two sets of timehands, the th_generation count
would always be too old to allow the captured PPS data to be used.
This commit is contained in:
Ian Lepore 2019-09-10 22:08:34 +00:00
parent e7da26b6ca
commit 192122bd62
Notes: svn2git 2020-12-20 02:59:44 +00:00 
svn path=/head/; revision=352196
1 changed files with 29 additions and 43 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

72
sys/arm/ti/am335x/am335x_dmtpps.c
View File

@ -52,7 +52,6 @@ __FBSDID("$FreeBSD$");
#include <sys/malloc.h>
#include <sys/mutex.h>
#include <sys/rman.h>
#include <sys/taskqueue.h>
#include <sys/timepps.h>
#include <sys/timetc.h>
#include <machine/bus.h>
@ -79,7 +78,6 @@ struct dmtpps_softc {
uint32_t tclr; /* Cached TCLR register. */
struct timecounter tc;
int pps_curmode; /* Edge mode now set in hw. */
struct task pps_task; /* For pps_event handling. */
struct cdev * pps_cdev;
struct pps_state pps_state;
struct mtx pps_mtx;
@ -93,6 +91,7 @@ static struct ofw_compat_data compat_data[] = {
{"ti,am335x-timer-1ms", 1},
{NULL, 0},
};
SIMPLEBUS_PNP_INFO(compat_data);
/*
* A table relating pad names to the hardware timer number they can be mux'd to.
@ -285,48 +284,29 @@ dmtpps_poll(struct timecounter *tc)
* populates it from the current DMT_TCRR register) with the latched
* value from the TCAR1 register.
*
* There is no locking here, by design. pps_capture() writes into an
* area of struct pps_state which is read only by pps_event(). The
* synchronization of access to that area is temporal rather than
* interlock based... we write in this routine and trigger the task that
* will read the data, so no simultaneous access can occur.
*
* Note that we don't have the TCAR interrupt enabled, but the hardware
* still provides the status bits in the "RAW" status register even when
* they're masked from generating an irq. However, when clearing the
* TCAR status to re-arm the capture for the next second, we have to
* write to the IRQ status register, not the RAW register. Quirky.
*
* We do not need to hold a lock while capturing the pps data, because
* it is captured into an area of the pps_state struct which is read
* only by pps_event(). We do need to hold a lock while calling
* pps_event(), because it manipulates data which is also accessed from
* the ioctl(2) context by userland processes.
*/
if (DMTIMER_READ4(sc, DMT_IRQSTATUS_RAW) & DMT_IRQ_TCAR) {
pps_capture(&sc->pps_state);
sc->pps_state.capcount = DMTIMER_READ4(sc, DMT_TCAR1);
DMTIMER_WRITE4(sc, DMT_IRQSTATUS, DMT_IRQ_TCAR);
taskqueue_enqueue(taskqueue_fast, &sc->pps_task);
mtx_lock_spin(&sc->pps_mtx);
pps_event(&sc->pps_state, PPS_CAPTUREASSERT);
mtx_unlock_spin(&sc->pps_mtx);
}
}
static void
dmtpps_event(void *arg, int pending)
{
struct dmtpps_softc *sc;
sc = arg;
/* This is the task function that gets enqueued by poll_pps. Once the
* time has been captured by the timecounter polling code which runs in
* primary interrupt context, the remaining (more expensive) work to
* process the event is done later in a threaded context.
*
* Here there is an interlock that protects the event data in struct
* pps_state. That data can be accessed at any time from userland via
* ioctl() calls so we must ensure that there is no read access to
* partially updated data while pps_event() does its work.
*/
mtx_lock(&sc->pps_mtx);
pps_event(&sc->pps_state, PPS_CAPTUREASSERT);
mtx_unlock(&sc->pps_mtx);
}
static int
dmtpps_open(struct cdev *dev, int flags, int fmt,
struct thread *td)
@ -374,9 +354,9 @@ dmtpps_ioctl(struct cdev *dev, u_long cmd, caddr_t data,
sc = dev->si_drv1;
/* Let the kernel do the heavy lifting for ioctl. */
mtx_lock(&sc->pps_mtx);
mtx_lock_spin(&sc->pps_mtx);
err = pps_ioctl(cmd, data, &sc->pps_state);
mtx_unlock(&sc->pps_mtx);
mtx_unlock_spin(&sc->pps_mtx);
if (err != 0)
return (err);
@ -436,6 +416,7 @@ static int
dmtpps_attach(device_t dev)
{
struct dmtpps_softc *sc;
struct make_dev_args mda;
clk_ident_t timer_id;
int err, sysclk_freq;
@ -502,22 +483,27 @@ dmtpps_attach(device_t dev)
* now, just say we can only capture assert events (the positive-going
* edge of the pulse).
*/
mtx_init(&sc->pps_mtx, "dmtpps", NULL, MTX_DEF);
mtx_init(&sc->pps_mtx, "dmtpps", NULL, MTX_SPIN);
sc->pps_state.flags = PPSFLAG_MTX_SPIN;
sc->pps_state.ppscap = PPS_CAPTUREASSERT;
sc->pps_state.driver_abi = PPS_ABI_VERSION;
sc->pps_state.driver_mtx = &sc->pps_mtx;
pps_init_abi(&sc->pps_state);
/*
* Init the task that does deferred pps_event() processing after
* the polling routine has captured a pps pulse time.
*/
TASK_INIT(&sc->pps_task, 0, dmtpps_event, sc);
/* Create the PPS cdev. */
sc->pps_cdev = make_dev(&dmtpps_cdevsw, 0, UID_ROOT, GID_WHEEL, 0600,
PPS_CDEV_NAME);
sc->pps_cdev->si_drv1 = sc;
make_dev_args_init(&mda);
mda.mda_flags = MAKEDEV_WAITOK;
mda.mda_devsw = &dmtpps_cdevsw;
mda.mda_cr = NULL;
mda.mda_uid = UID_ROOT;
mda.mda_gid = GID_WHEEL;
mda.mda_mode = 0600;
mda.mda_unit = device_get_unit(dev);
mda.mda_si_drv1 = sc;
if ((err = make_dev_s(&mda, &sc->pps_cdev, PPS_CDEV_NAME)) != 0) {
device_printf(dev, "Failed to create cdev %s\n", PPS_CDEV_NAME);
return (err);
}
if (bootverbose)
device_printf(sc->dev, "Using %s for PPS device /dev/%s\n",