Re: [PATCH v2] serial: stm32: optimize spin lock usage
Hi Johan, On Fri, Apr 16, 2021 at 4:35 PM Johan Hovold wrote: > > On Tue, Apr 13, 2021 at 07:44:39AM +0800, dillon min wrote: > > Hi Johan, Erwan > > > > It seems still a bit of a problem in the current version, not deadlock > > but access register at the same time. > > > > For driver , we should consider it running under smp, let's think > > about it for this case: > > > > static void stm32_usart_console_write(struct console *co, const char *s, > > unsigned int cnt) > > { > > . > > local_irq_save(flags); > > if (port->sysrq) > > locked = 0; > > . > > access register cr1, tdr, isr > > . > > > > local_irq_restore(flags); > > } > > > > if port->sysrq is 1, stm32_usart_console_write() just disable local > > irq response by local_irq_save(), at the time of access register cr1, > > tdr, isr. an TXE interrupt raised, for other cores(I know stm32 > > mpu/mcu do not have multi cores, just assume it has), it still has a > > chance to handle interrupt. Then there is no lock to protect the uart > > register. > > Right, the sysrq handling is a bit of a hack. > > > changes to below, should be more safe: > > > > . > > if (port->sysrq || oops_in_progress) > > locked = spin_trylock_irqsave(>lock, flags); > > Except that the lock debugging code would detect the attempt at > recursive locking here and complain loudly on UP. > > If you really want to fix this, we have uart_unlock_and_check_sysrq() > which can be used to defer sysrq processing until the interrupt handler > has released the lock. Great, uart_unlock_and_check_sysrq() is fit to fix this. you mean make the flow like below: stm32_usart_threaded_interrupt() spin_lock(>lock); uart_unlock_and_check_sysrq(port, flags); ... uart_prepare_sysrq_char(); printk(); stm32_usart_console_write(); locked = spin_trylock_irqsave(>lock); //only handle oops, normal case If so, I will submit v3 as you suggested. thanks. Best regards. Dillon, > > > else > > spin_lock_irqsave(>lock, flags); > > > > > > > > if (locked) > > spin_unlock_irqrestore(>lock, flags); > > Johan
Re: [PATCH v2] serial: stm32: optimize spin lock usage
[ Please avoid top-posting. ] On Thu, Apr 15, 2021 at 07:09:14PM +0200, Erwan LE RAY wrote: > Hi Dillon, > > STM32MP151 is mono-core, but both STM32MP153 and STM32MP157 are > dual-core (see > https://www.st.com/content/st_com/en/products/microcontrollers-microprocessors/stm32-arm-cortex-mpus.html). > So your point is fully relevant, thanks. > > ST already fixed the same issue in st-asc.c driver in the past (see > ef49ffd8), because a systematic deadlock was detected with RT kernel. That's not the same issue. The above mentioned commit fixed an issue on *RT* where local_irq_save() should be avoided. > You proposed a first implementation in your patch, and a second one in > the discussion. It seems that your initial proposal (ie your V2 patch) > is the most standard one (implemented in 6 drivers). The second > implementation is implemented by only 1 company. > > It looks that the solution is to avoid locking in the sysrq case and > trylock in the oops_in_progress case (see detailed analysis in > 677fe555cbfb1). > > So your initial patch looks to the right proposal, but it would be safer > if Greg could confirm it. That would only fix the RT issue (and by making the sysrq one slightly worse). Using uart_unlock_and_check_sysrq() would address both issues. Johan
Re: [PATCH v2] serial: stm32: optimize spin lock usage
On Tue, Apr 13, 2021 at 07:44:39AM +0800, dillon min wrote: > Hi Johan, Erwan > > It seems still a bit of a problem in the current version, not deadlock > but access register at the same time. > > For driver , we should consider it running under smp, let's think > about it for this case: > > static void stm32_usart_console_write(struct console *co, const char *s, > unsigned int cnt) > { > . > local_irq_save(flags); > if (port->sysrq) > locked = 0; > . > access register cr1, tdr, isr > . > > local_irq_restore(flags); > } > > if port->sysrq is 1, stm32_usart_console_write() just disable local > irq response by local_irq_save(), at the time of access register cr1, > tdr, isr. an TXE interrupt raised, for other cores(I know stm32 > mpu/mcu do not have multi cores, just assume it has), it still has a > chance to handle interrupt. Then there is no lock to protect the uart > register. Right, the sysrq handling is a bit of a hack. > changes to below, should be more safe: > > . > if (port->sysrq || oops_in_progress) > locked = spin_trylock_irqsave(>lock, flags); Except that the lock debugging code would detect the attempt at recursive locking here and complain loudly on UP. If you really want to fix this, we have uart_unlock_and_check_sysrq() which can be used to defer sysrq processing until the interrupt handler has released the lock. > else > spin_lock_irqsave(>lock, flags); > > > > if (locked) > spin_unlock_irqrestore(>lock, flags); Johan
Re: [PATCH v2] serial: stm32: optimize spin lock usage
Hi Erwan, Erwan LE RAY 于2021年4月16日周五 上午1:10写道: > > Hi Dillon, > > STM32MP151 is mono-core, but both STM32MP153 and STM32MP157 are > dual-core (see > https://www.st.com/content/st_com/en/products/microcontrollers-microprocessors/stm32-arm-cortex-mpus.html). > So your point is fully relevant, thanks. Thanks. > > ST already fixed the same issue in st-asc.c driver in the past (see > ef49ffd8), because a systematic deadlock was detected with RT kernel. > > You proposed a first implementation in your patch, and a second one in > the discussion. It seems that your initial proposal (ie your V2 patch) > is the most standard one (implemented in 6 drivers). The second > implementation is implemented by only 1 company. > > It looks that the solution is to avoid locking in the sysrq case and > trylock in the oops_in_progress case (see detailed analysis in > 677fe555cbfb1). Thanks for the detail information. the V2 patch didn't cover this case: stm32_usart_threaded_interrupt() spin_lock(>lock); ... uart_handle_sysrq_char(); sysrq_function(); printk(); stm32_usart_console_write(); locked = spin_trylock_irqsave(>lock); //better than no lock(locked = 0) if other uart interrupt coming at this point Find a same solution on fsl_lpuart.c commit abf1e0a98083fd0a1069ce68ad8c92bfb97a57db Thanks. Best regards Dillon > > So your initial patch looks to the right proposal, but it would be safer > if Greg could confirm it. > > BR, Erwan. > > > On 4/13/21 1:44 AM, dillon min wrote: > > Hi Johan, Erwan > > > > It seems still a bit of a problem in the current version, not deadlock > > but access register at the same time. > > > > For driver , we should consider it running under smp, let's think > > about it for this case: > > > > static void stm32_usart_console_write(struct console *co, const char *s, > >unsigned int cnt) > > { > > . > > local_irq_save(flags); > > if (port->sysrq) > > locked = 0; > > . > > access register cr1, tdr, isr > > . > > > > local_irq_restore(flags); > > } > > > > if port->sysrq is 1, stm32_usart_console_write() just disable local > > irq response by local_irq_save(), at the time of access register cr1, > > tdr, isr. an TXE interrupt raised, for other cores(I know stm32 > > mpu/mcu do not have multi cores, just assume it has), it still has a > > chance to handle interrupt. Then there is no lock to protect the uart > > register. > > > > changes to below, should be more safe: > > > > . > > if (port->sysrq || oops_in_progress) > >locked = spin_trylock_irqsave(>lock, flags); > > else > >spin_lock_irqsave(>lock, flags); > > > > > > > > if (locked) > > spin_unlock_irqrestore(>lock, flags); > > > > For current stm32 soc, it shouldn't happen. just a reminder for future. > > > > Thanks. > > > > Dillon > > > > On Mon, Apr 12, 2021 at 10:04 PM dillon min wrote: > >> > >> Hi Johan, > >> > >> Yes, there is no deadlock. my fault. > >> I forget the local_irq_save() plus spin_lock() is spin_lock_irqsave(). > >> > >> Thanks for your review. please ignore this patch. > >> > >> Best regards > >> > >> Dillon > >> > >> On Mon, Apr 12, 2021 at 9:08 PM Johan Hovold wrote: > >>> > >>> On Mon, Apr 12, 2021 at 05:31:38PM +0800, dillon.min...@gmail.com wrote: > From: dillon min > > To avoid potential deadlock in spin_lock usage, use spin_lock_irqsave, > spin_trylock_irqsave(), spin_unlock_irqrestore() in process context. > >>> > >>> This doesn't make much sense as console_write can be called in any > >>> context. And where's the deadlock you claim to be fixing here? > >>> > remove unused local_irq_save/restore call. > > Cc: Alexandre Torgue > Cc: Maxime Coquelin > Cc: Gerald Baeza > Cc: Erwan Le Ray > Reported-by: kernel test robot > Signed-off-by: dillon min > --- > v2: remove unused code from stm32_usart_threaded_interrupt() according > from > Greg's review. > > drivers/tty/serial/stm32-usart.c | 8 +++- > 1 file changed, 3 insertions(+), 5 deletions(-) > > diff --git a/drivers/tty/serial/stm32-usart.c > b/drivers/tty/serial/stm32-usart.c > index b3675cf25a69..b1ba5e36e36e 100644 > --- a/drivers/tty/serial/stm32-usart.c > +++ b/drivers/tty/serial/stm32-usart.c > @@ -1354,13 +1354,12 @@ static void stm32_usart_console_write(struct > console *co, const char *s, > u32 old_cr1, new_cr1; > int locked = 1; > > - local_irq_save(flags); > if (port->sysrq) > locked = 0; > else if (oops_in_progress) > - locked = spin_trylock(>lock); > + locked = spin_trylock_irqsave(>lock, flags); > else > -
Re: [PATCH v2] serial: stm32: optimize spin lock usage
Hi Dillon, STM32MP151 is mono-core, but both STM32MP153 and STM32MP157 are dual-core (see https://www.st.com/content/st_com/en/products/microcontrollers-microprocessors/stm32-arm-cortex-mpus.html). So your point is fully relevant, thanks. ST already fixed the same issue in st-asc.c driver in the past (see ef49ffd8), because a systematic deadlock was detected with RT kernel. You proposed a first implementation in your patch, and a second one in the discussion. It seems that your initial proposal (ie your V2 patch) is the most standard one (implemented in 6 drivers). The second implementation is implemented by only 1 company. It looks that the solution is to avoid locking in the sysrq case and trylock in the oops_in_progress case (see detailed analysis in 677fe555cbfb1). So your initial patch looks to the right proposal, but it would be safer if Greg could confirm it. BR, Erwan. On 4/13/21 1:44 AM, dillon min wrote: Hi Johan, Erwan It seems still a bit of a problem in the current version, not deadlock but access register at the same time. For driver , we should consider it running under smp, let's think about it for this case: static void stm32_usart_console_write(struct console *co, const char *s, unsigned int cnt) { . local_irq_save(flags); if (port->sysrq) locked = 0; . access register cr1, tdr, isr . local_irq_restore(flags); } if port->sysrq is 1, stm32_usart_console_write() just disable local irq response by local_irq_save(), at the time of access register cr1, tdr, isr. an TXE interrupt raised, for other cores(I know stm32 mpu/mcu do not have multi cores, just assume it has), it still has a chance to handle interrupt. Then there is no lock to protect the uart register. changes to below, should be more safe: . if (port->sysrq || oops_in_progress) locked = spin_trylock_irqsave(>lock, flags); else spin_lock_irqsave(>lock, flags); if (locked) spin_unlock_irqrestore(>lock, flags); For current stm32 soc, it shouldn't happen. just a reminder for future. Thanks. Dillon On Mon, Apr 12, 2021 at 10:04 PM dillon min wrote: Hi Johan, Yes, there is no deadlock. my fault. I forget the local_irq_save() plus spin_lock() is spin_lock_irqsave(). Thanks for your review. please ignore this patch. Best regards Dillon On Mon, Apr 12, 2021 at 9:08 PM Johan Hovold wrote: On Mon, Apr 12, 2021 at 05:31:38PM +0800, dillon.min...@gmail.com wrote: From: dillon min To avoid potential deadlock in spin_lock usage, use spin_lock_irqsave, spin_trylock_irqsave(), spin_unlock_irqrestore() in process context. This doesn't make much sense as console_write can be called in any context. And where's the deadlock you claim to be fixing here? remove unused local_irq_save/restore call. Cc: Alexandre Torgue Cc: Maxime Coquelin Cc: Gerald Baeza Cc: Erwan Le Ray Reported-by: kernel test robot Signed-off-by: dillon min --- v2: remove unused code from stm32_usart_threaded_interrupt() according from Greg's review. drivers/tty/serial/stm32-usart.c | 8 +++- 1 file changed, 3 insertions(+), 5 deletions(-) diff --git a/drivers/tty/serial/stm32-usart.c b/drivers/tty/serial/stm32-usart.c index b3675cf25a69..b1ba5e36e36e 100644 --- a/drivers/tty/serial/stm32-usart.c +++ b/drivers/tty/serial/stm32-usart.c @@ -1354,13 +1354,12 @@ static void stm32_usart_console_write(struct console *co, const char *s, u32 old_cr1, new_cr1; int locked = 1; - local_irq_save(flags); if (port->sysrq) locked = 0; else if (oops_in_progress) - locked = spin_trylock(>lock); + locked = spin_trylock_irqsave(>lock, flags); else - spin_lock(>lock); + spin_lock_irqsave(>lock, flags); /* Save and disable interrupts, enable the transmitter */ old_cr1 = readl_relaxed(port->membase + ofs->cr1); @@ -1374,8 +1373,7 @@ static void stm32_usart_console_write(struct console *co, const char *s, writel_relaxed(old_cr1, port->membase + ofs->cr1); if (locked) - spin_unlock(>lock); - local_irq_restore(flags); + spin_unlock_irqrestore(>lock, flags); } static int stm32_usart_console_setup(struct console *co, char *options) Johan
Re: [PATCH v2] serial: stm32: optimize spin lock usage
Hi Johan, Erwan It seems still a bit of a problem in the current version, not deadlock but access register at the same time. For driver , we should consider it running under smp, let's think about it for this case: static void stm32_usart_console_write(struct console *co, const char *s, unsigned int cnt) { . local_irq_save(flags); if (port->sysrq) locked = 0; . access register cr1, tdr, isr . local_irq_restore(flags); } if port->sysrq is 1, stm32_usart_console_write() just disable local irq response by local_irq_save(), at the time of access register cr1, tdr, isr. an TXE interrupt raised, for other cores(I know stm32 mpu/mcu do not have multi cores, just assume it has), it still has a chance to handle interrupt. Then there is no lock to protect the uart register. changes to below, should be more safe: . if (port->sysrq || oops_in_progress) locked = spin_trylock_irqsave(>lock, flags); else spin_lock_irqsave(>lock, flags); if (locked) spin_unlock_irqrestore(>lock, flags); For current stm32 soc, it shouldn't happen. just a reminder for future. Thanks. Dillon On Mon, Apr 12, 2021 at 10:04 PM dillon min wrote: > > Hi Johan, > > Yes, there is no deadlock. my fault. > I forget the local_irq_save() plus spin_lock() is spin_lock_irqsave(). > > Thanks for your review. please ignore this patch. > > Best regards > > Dillon > > On Mon, Apr 12, 2021 at 9:08 PM Johan Hovold wrote: > > > > On Mon, Apr 12, 2021 at 05:31:38PM +0800, dillon.min...@gmail.com wrote: > > > From: dillon min > > > > > > To avoid potential deadlock in spin_lock usage, use spin_lock_irqsave, > > > spin_trylock_irqsave(), spin_unlock_irqrestore() in process context. > > > > This doesn't make much sense as console_write can be called in any > > context. And where's the deadlock you claim to be fixing here? > > > > > remove unused local_irq_save/restore call. > > > > > > Cc: Alexandre Torgue > > > Cc: Maxime Coquelin > > > Cc: Gerald Baeza > > > Cc: Erwan Le Ray > > > Reported-by: kernel test robot > > > Signed-off-by: dillon min > > > --- > > > v2: remove unused code from stm32_usart_threaded_interrupt() according > > > from > > > Greg's review. > > > > > > drivers/tty/serial/stm32-usart.c | 8 +++- > > > 1 file changed, 3 insertions(+), 5 deletions(-) > > > > > > diff --git a/drivers/tty/serial/stm32-usart.c > > > b/drivers/tty/serial/stm32-usart.c > > > index b3675cf25a69..b1ba5e36e36e 100644 > > > --- a/drivers/tty/serial/stm32-usart.c > > > +++ b/drivers/tty/serial/stm32-usart.c > > > @@ -1354,13 +1354,12 @@ static void stm32_usart_console_write(struct > > > console *co, const char *s, > > > u32 old_cr1, new_cr1; > > > int locked = 1; > > > > > > - local_irq_save(flags); > > > if (port->sysrq) > > > locked = 0; > > > else if (oops_in_progress) > > > - locked = spin_trylock(>lock); > > > + locked = spin_trylock_irqsave(>lock, flags); > > > else > > > - spin_lock(>lock); > > > + spin_lock_irqsave(>lock, flags); > > > > > > /* Save and disable interrupts, enable the transmitter */ > > > old_cr1 = readl_relaxed(port->membase + ofs->cr1); > > > @@ -1374,8 +1373,7 @@ static void stm32_usart_console_write(struct > > > console *co, const char *s, > > > writel_relaxed(old_cr1, port->membase + ofs->cr1); > > > > > > if (locked) > > > - spin_unlock(>lock); > > > - local_irq_restore(flags); > > > + spin_unlock_irqrestore(>lock, flags); > > > } > > > > > > static int stm32_usart_console_setup(struct console *co, char *options) > > > > Johan
Re: [PATCH v2] serial: stm32: optimize spin lock usage
Hi Johan, Yes, there is no deadlock. my fault. I forget the local_irq_save() plus spin_lock() is spin_lock_irqsave(). Thanks for your review. please ignore this patch. Best regards Dillon On Mon, Apr 12, 2021 at 9:08 PM Johan Hovold wrote: > > On Mon, Apr 12, 2021 at 05:31:38PM +0800, dillon.min...@gmail.com wrote: > > From: dillon min > > > > To avoid potential deadlock in spin_lock usage, use spin_lock_irqsave, > > spin_trylock_irqsave(), spin_unlock_irqrestore() in process context. > > This doesn't make much sense as console_write can be called in any > context. And where's the deadlock you claim to be fixing here? > > > remove unused local_irq_save/restore call. > > > > Cc: Alexandre Torgue > > Cc: Maxime Coquelin > > Cc: Gerald Baeza > > Cc: Erwan Le Ray > > Reported-by: kernel test robot > > Signed-off-by: dillon min > > --- > > v2: remove unused code from stm32_usart_threaded_interrupt() according from > > Greg's review. > > > > drivers/tty/serial/stm32-usart.c | 8 +++- > > 1 file changed, 3 insertions(+), 5 deletions(-) > > > > diff --git a/drivers/tty/serial/stm32-usart.c > > b/drivers/tty/serial/stm32-usart.c > > index b3675cf25a69..b1ba5e36e36e 100644 > > --- a/drivers/tty/serial/stm32-usart.c > > +++ b/drivers/tty/serial/stm32-usart.c > > @@ -1354,13 +1354,12 @@ static void stm32_usart_console_write(struct > > console *co, const char *s, > > u32 old_cr1, new_cr1; > > int locked = 1; > > > > - local_irq_save(flags); > > if (port->sysrq) > > locked = 0; > > else if (oops_in_progress) > > - locked = spin_trylock(>lock); > > + locked = spin_trylock_irqsave(>lock, flags); > > else > > - spin_lock(>lock); > > + spin_lock_irqsave(>lock, flags); > > > > /* Save and disable interrupts, enable the transmitter */ > > old_cr1 = readl_relaxed(port->membase + ofs->cr1); > > @@ -1374,8 +1373,7 @@ static void stm32_usart_console_write(struct console > > *co, const char *s, > > writel_relaxed(old_cr1, port->membase + ofs->cr1); > > > > if (locked) > > - spin_unlock(>lock); > > - local_irq_restore(flags); > > + spin_unlock_irqrestore(>lock, flags); > > } > > > > static int stm32_usart_console_setup(struct console *co, char *options) > > Johan
Re: [PATCH v2] serial: stm32: optimize spin lock usage
On Mon, Apr 12, 2021 at 05:31:38PM +0800, dillon.min...@gmail.com wrote: > From: dillon min > > To avoid potential deadlock in spin_lock usage, use spin_lock_irqsave, > spin_trylock_irqsave(), spin_unlock_irqrestore() in process context. This doesn't make much sense as console_write can be called in any context. And where's the deadlock you claim to be fixing here? > remove unused local_irq_save/restore call. > > Cc: Alexandre Torgue > Cc: Maxime Coquelin > Cc: Gerald Baeza > Cc: Erwan Le Ray > Reported-by: kernel test robot > Signed-off-by: dillon min > --- > v2: remove unused code from stm32_usart_threaded_interrupt() according from > Greg's review. > > drivers/tty/serial/stm32-usart.c | 8 +++- > 1 file changed, 3 insertions(+), 5 deletions(-) > > diff --git a/drivers/tty/serial/stm32-usart.c > b/drivers/tty/serial/stm32-usart.c > index b3675cf25a69..b1ba5e36e36e 100644 > --- a/drivers/tty/serial/stm32-usart.c > +++ b/drivers/tty/serial/stm32-usart.c > @@ -1354,13 +1354,12 @@ static void stm32_usart_console_write(struct console > *co, const char *s, > u32 old_cr1, new_cr1; > int locked = 1; > > - local_irq_save(flags); > if (port->sysrq) > locked = 0; > else if (oops_in_progress) > - locked = spin_trylock(>lock); > + locked = spin_trylock_irqsave(>lock, flags); > else > - spin_lock(>lock); > + spin_lock_irqsave(>lock, flags); > > /* Save and disable interrupts, enable the transmitter */ > old_cr1 = readl_relaxed(port->membase + ofs->cr1); > @@ -1374,8 +1373,7 @@ static void stm32_usart_console_write(struct console > *co, const char *s, > writel_relaxed(old_cr1, port->membase + ofs->cr1); > > if (locked) > - spin_unlock(>lock); > - local_irq_restore(flags); > + spin_unlock_irqrestore(>lock, flags); > } > > static int stm32_usart_console_setup(struct console *co, char *options) Johan