| Message ID | 2f3ba828505cb3e8f9dc8a7b6c5a58a51a80cd90.1585445576.git.richardcochran@gmail.com |
|---|---|
| State | Accepted |
| Delegated to: | David Miller |
| Headers | show |
| Series | [net-next,1/1] ptp: Avoid deadlocks in the programmable pin code. | expand |
Hi Richard, On Sun, 29 Mar 2020 at 04:40, Richard Cochran <richardcochran@gmail.com> wrote: > > The PTP Hardware Clock (PHC) subsystem offers an API for configuring > programmable pins. User space sets or gets the settings using ioctls, > and drivers verify dialed settings via a callback. Drivers may also > query pin settings by calling the ptp_find_pin() method. > > Although the core subsystem protects concurrent access to the pin > settings, the implementation places illogical restrictions on how > drivers may call ptp_find_pin(). When enabling an auxiliary function > via the .enable(on=1) callback, drivers may invoke the pin finding > method, but when disabling with .enable(on=0) drivers are not > permitted to do so. With the exception of the mv88e6xxx, all of the > PHC drivers do respect this restriction, but still the locking pattern > is both confusing and unnecessary. > > This patch changes the locking implementation to allow PHC drivers to > freely call ptp_find_pin() from their .enable() and .verify() > callbacks. > > Signed-off-by: Richard Cochran <richardcochran@gmail.com> > Reported-by: Yangbo Lu <yangbo.lu@nxp.com> > --- I've tested this on top of Yangbo's patch, using this diff: drivers/ptp/ptp_ocelot.c | 49 +++++++++++++++++----------------------- 1 file changed, 21 insertions(+), 28 deletions(-) diff --git a/drivers/ptp/ptp_ocelot.c b/drivers/ptp/ptp_ocelot.c index 299928e8691d..7d35ba262278 100644 --- a/drivers/ptp/ptp_ocelot.c +++ b/drivers/ptp/ptp_ocelot.c @@ -183,11 +183,9 @@ static int ocelot_ptp_enable(struct ptp_clock_info *ptp, { struct ocelot *ocelot = container_of(ptp, struct ocelot, ptp_info); enum ocelot_ptp_pins ptp_pin; - struct timespec64 ts; unsigned long flags; int pin = -1; u32 val; - s64 ns; switch (rq->type) { case PTP_CLK_REQ_PEROUT: @@ -195,18 +193,6 @@ static int ocelot_ptp_enable(struct ptp_clock_info *ptp, if (rq->perout.flags) return -EOPNOTSUPP; - /* - * TODO: support disabling function - * When ptp_disable_pinfunc() is to disable function, - * it has already held pincfg_mux. - * However ptp_find_pin() in .enable() called also needs - * to hold pincfg_mux. - * This causes dead lock. So, just return for function - * disabling, and this needs fix-up. - */ - if (!on) - break; - pin = ptp_find_pin(ocelot->ptp_clock, PTP_PF_PEROUT, rq->perout.index); if (pin == 0) @@ -220,22 +206,29 @@ static int ocelot_ptp_enable(struct ptp_clock_info *ptp, else return -EINVAL; - ts.tv_sec = rq->perout.period.sec; - ts.tv_nsec = rq->perout.period.nsec; - ns = timespec64_to_ns(&ts); - ns = ns >> 1; - if (ns > 0x3fffffff || ns <= 0x6) - return -EINVAL; - spin_lock_irqsave(&ocelot->ptp_clock_lock, flags); - ocelot_write_rix(ocelot, ns, PTP_PIN_WF_LOW_PERIOD, ptp_pin); - ocelot_write_rix(ocelot, ns, PTP_PIN_WF_HIGH_PERIOD, ptp_pin); - - val = PTP_PIN_CFG_ACTION(PTP_PIN_ACTION_CLOCK); - ocelot_write_rix(ocelot, val, PTP_PIN_CFG, ptp_pin); + if (on) { + struct timespec64 ts = { + .tv_sec = rq->perout.period.sec, + .tv_nsec = rq->perout.period.nsec, + }; + s64 ns = timespec64_to_ns(&ts) >> 1; + + if (ns > 0x3fffffff || ns <= 0x6) + return -EINVAL; + + ocelot_write_rix(ocelot, ns, PTP_PIN_WF_LOW_PERIOD, + ptp_pin); + ocelot_write_rix(ocelot, ns, PTP_PIN_WF_HIGH_PERIOD, + ptp_pin); + + val = PTP_PIN_CFG_ACTION(PTP_PIN_ACTION_CLOCK); + ocelot_write_rix(ocelot, val, PTP_PIN_CFG, ptp_pin); + } else { + val = PTP_PIN_CFG_ACTION(PTP_PIN_ACTION_IDLE); + ocelot_write_rix(ocelot, val, PTP_PIN_CFG, ptp_pin); + } spin_unlock_irqrestore(&ocelot->ptp_clock_lock, flags); - dev_warn(ocelot->dev, - "Starting periodic signal now! (absolute start time not supported)\n"); break; default: return -EOPNOTSUPP;
On Sun, Mar 29, 2020 at 03:56:18PM +0300, Vladimir Oltean wrote: > > @@ -175,7 +175,10 @@ long ptp_ioctl(struct posix_clock *pc, unsigned int cmd, unsigned long arg) > > } > > req.type = PTP_CLK_REQ_EXTTS; > > enable = req.extts.flags & PTP_ENABLE_FEATURE ? 1 : 0; > > + if (mutex_lock_interruptible(&ptp->pincfg_mux)) > > + return -ERESTARTSYS; > > Is there any reason why you're not just propagating the return value > of mutex_lock_interruptible? Yes, this return code lets the system call be able to be restarted after interruption by a signal. Thanks, Richard
From: Richard Cochran <richardcochran@gmail.com> Date: Sat, 28 Mar 2020 18:35:30 -0700 > The PTP Hardware Clock (PHC) subsystem offers an API for configuring > programmable pins. User space sets or gets the settings using ioctls, > and drivers verify dialed settings via a callback. Drivers may also > query pin settings by calling the ptp_find_pin() method. > > Although the core subsystem protects concurrent access to the pin > settings, the implementation places illogical restrictions on how > drivers may call ptp_find_pin(). When enabling an auxiliary function > via the .enable(on=1) callback, drivers may invoke the pin finding > method, but when disabling with .enable(on=0) drivers are not > permitted to do so. With the exception of the mv88e6xxx, all of the > PHC drivers do respect this restriction, but still the locking pattern > is both confusing and unnecessary. > > This patch changes the locking implementation to allow PHC drivers to > freely call ptp_find_pin() from their .enable() and .verify() > callbacks. > > Signed-off-by: Richard Cochran <richardcochran@gmail.com> > Reported-by: Yangbo Lu <yangbo.lu@nxp.com> Applied.
diff --git a/drivers/net/phy/dp83640.c b/drivers/net/phy/dp83640.c index ac72a324fcd1..415c27310982 100644 --- a/drivers/net/phy/dp83640.c +++ b/drivers/net/phy/dp83640.c @@ -628,7 +628,7 @@ static void recalibrate(struct dp83640_clock *clock) u16 cal_gpio, cfg0, evnt, ptp_trig, trigger, val; trigger = CAL_TRIGGER; - cal_gpio = 1 + ptp_find_pin(clock->ptp_clock, PTP_PF_PHYSYNC, 0); + cal_gpio = 1 + ptp_find_pin_unlocked(clock->ptp_clock, PTP_PF_PHYSYNC, 0); if (cal_gpio < 1) { pr_err("PHY calibration pin not available - PHY is not calibrated."); return; diff --git a/drivers/ptp/ptp_chardev.c b/drivers/ptp/ptp_chardev.c index 9d72ab593f13..93d574faf1fe 100644 --- a/drivers/ptp/ptp_chardev.c +++ b/drivers/ptp/ptp_chardev.c @@ -175,7 +175,10 @@ long ptp_ioctl(struct posix_clock *pc, unsigned int cmd, unsigned long arg) } req.type = PTP_CLK_REQ_EXTTS; enable = req.extts.flags & PTP_ENABLE_FEATURE ? 1 : 0; + if (mutex_lock_interruptible(&ptp->pincfg_mux)) + return -ERESTARTSYS; err = ops->enable(ops, &req, enable); + mutex_unlock(&ptp->pincfg_mux); break; case PTP_PEROUT_REQUEST: @@ -206,7 +209,10 @@ long ptp_ioctl(struct posix_clock *pc, unsigned int cmd, unsigned long arg) } req.type = PTP_CLK_REQ_PEROUT; enable = req.perout.period.sec || req.perout.period.nsec; + if (mutex_lock_interruptible(&ptp->pincfg_mux)) + return -ERESTARTSYS; err = ops->enable(ops, &req, enable); + mutex_unlock(&ptp->pincfg_mux); break; case PTP_ENABLE_PPS: @@ -217,7 +223,10 @@ long ptp_ioctl(struct posix_clock *pc, unsigned int cmd, unsigned long arg) return -EPERM; req.type = PTP_CLK_REQ_PPS; enable = arg ? 1 : 0; + if (mutex_lock_interruptible(&ptp->pincfg_mux)) + return -ERESTARTSYS; err = ops->enable(ops, &req, enable); + mutex_unlock(&ptp->pincfg_mux); break; case PTP_SYS_OFFSET_PRECISE: diff --git a/drivers/ptp/ptp_clock.c b/drivers/ptp/ptp_clock.c index ac1f2bf9e888..acabbe72e55e 100644 --- a/drivers/ptp/ptp_clock.c +++ b/drivers/ptp/ptp_clock.c @@ -348,7 +348,6 @@ int ptp_find_pin(struct ptp_clock *ptp, struct ptp_pin_desc *pin = NULL; int i; - mutex_lock(&ptp->pincfg_mux); for (i = 0; i < ptp->info->n_pins; i++) { if (ptp->info->pin_config[i].func == func && ptp->info->pin_config[i].chan == chan) { @@ -356,12 +355,26 @@ int ptp_find_pin(struct ptp_clock *ptp, break; } } - mutex_unlock(&ptp->pincfg_mux); return pin ? i : -1; } EXPORT_SYMBOL(ptp_find_pin); +int ptp_find_pin_unlocked(struct ptp_clock *ptp, + enum ptp_pin_function func, unsigned int chan) +{ + int result; + + mutex_lock(&ptp->pincfg_mux); + + result = ptp_find_pin(ptp, func, chan); + + mutex_unlock(&ptp->pincfg_mux); + + return result; +} +EXPORT_SYMBOL(ptp_find_pin_unlocked); + int ptp_schedule_worker(struct ptp_clock *ptp, unsigned long delay) { return kthread_mod_delayed_work(ptp->kworker, &ptp->aux_work, delay); diff --git a/include/linux/ptp_clock_kernel.h b/include/linux/ptp_clock_kernel.h index c64a1ef87240..114807e7abdd 100644 --- a/include/linux/ptp_clock_kernel.h +++ b/include/linux/ptp_clock_kernel.h @@ -223,6 +223,12 @@ extern s32 scaled_ppm_to_ppb(long ppm); /** * ptp_find_pin() - obtain the pin index of a given auxiliary function * + * The caller must hold ptp_clock::pincfg_mux. Drivers do not have + * access to that mutex as ptp_clock is an opaque type. However, the + * core code acquires the mutex before invoking the driver's + * ptp_clock_info::enable() callback, and so drivers may call this + * function from that context. + * * @ptp: The clock obtained from ptp_clock_register(). * @func: One of the ptp_pin_function enumerated values. * @chan: The particular functional channel to find. @@ -233,6 +239,19 @@ extern s32 scaled_ppm_to_ppb(long ppm); int ptp_find_pin(struct ptp_clock *ptp, enum ptp_pin_function func, unsigned int chan); +/** + * ptp_find_pin_unlocked() - wrapper for ptp_find_pin() + * + * This function aquires the ptp_clock::pincfg_mux mutex before + * invoking ptp_find_pin(). Instead of using this function, drivers + * should most likely call ptp_find_pin() directly from their + * ptp_clock_info::enable() method. + * + */ + +int ptp_find_pin_unlocked(struct ptp_clock *ptp, + enum ptp_pin_function func, unsigned int chan); + /** * ptp_schedule_worker() - schedule ptp auxiliary work *
The PTP Hardware Clock (PHC) subsystem offers an API for configuring programmable pins. User space sets or gets the settings using ioctls, and drivers verify dialed settings via a callback. Drivers may also query pin settings by calling the ptp_find_pin() method. Although the core subsystem protects concurrent access to the pin settings, the implementation places illogical restrictions on how drivers may call ptp_find_pin(). When enabling an auxiliary function via the .enable(on=1) callback, drivers may invoke the pin finding method, but when disabling with .enable(on=0) drivers are not permitted to do so. With the exception of the mv88e6xxx, all of the PHC drivers do respect this restriction, but still the locking pattern is both confusing and unnecessary. This patch changes the locking implementation to allow PHC drivers to freely call ptp_find_pin() from their .enable() and .verify() callbacks. Signed-off-by: Richard Cochran <richardcochran@gmail.com> Reported-by: Yangbo Lu <yangbo.lu@nxp.com> --- drivers/net/phy/dp83640.c | 2 +- drivers/ptp/ptp_chardev.c | 9 +++++++++ drivers/ptp/ptp_clock.c | 17 +++++++++++++++-- include/linux/ptp_clock_kernel.h | 19 +++++++++++++++++++ 4 files changed, 44 insertions(+), 3 deletions(-)