Re: [PATCH v2 2/2] tty: serial: msm_serial: Fix deadlock caused by recursive output

From: Leo Yan
Date: Fri Dec 27 2019 - 00:52:46 EST


Hi Jeffrey,

On Mon, Dec 16, 2019 at 11:49:52AM -0700, Jeffrey Hugo wrote:
> On Wed, Dec 4, 2019 at 9:13 AM Leo Yan <leo.yan@xxxxxxxxxx> wrote:
> >
> > On Tue, Dec 03, 2019 at 03:42:31PM -0700, Jeffrey Hugo wrote:
> >
> > [...]
> >
> > > > > > This patch fixes the deadlock issue for recursive output; it adds a
> > > > > > variable 'curr_user' to indicate the uart port is used by which CPU, if
> > > > > > the CPU has acquired spinlock and wants to execute recursive output,
> > > > > > it will directly bail out. Here we don't choose to avoid locking and
> > > > > > print out log, the reason is in this case we don't want to reset the
> > > > > > uart port with function msm_reset_dm_count(); otherwise it can introduce
> > > > > > confliction with other flows and results in uart port malfunction and
> > > > > > later cannot output anymore.
> > > > >
> > > > > Is this not fixable? Sure, fixing the deadlock is an improvement, but
> > > > > dropping logs (particularly a memory warning like in your example)
> > > > > seems undesirable.
> > > >
> > > > Thanks a lot for your reviewing, Jeffrey.
> > > >
> > > > Agreed with you for the concern.
> > > >
> > > > To be honest, I am not familiar with the msm uart driver, so have no
> > > > confidence which is the best way for uart port operations. I can
> > > > think out one possible fixing is shown in below, if detects the lock
> > > > is not acquired then it will force to reset UART port before exit the
> > > > function __msm_console_write().
> > > >
> > > > This approach is not tested yet and it looks too arbitrary; I will
> > > > give a try for it. At the meantime, welcome any insight suggestion
> > > > with proper register operations.
> > >
> > > According to the documentation, NCF_TX is only needed for SW transmit
> > > mode, where software is directly puttting characters in the fifo. Its
> > > not needed for BAM mode. According to your example, recursive console
> > > printing will only happen in BAM mode, and not in SW mode. Perhaps if
> > > we put the NCF_TX uses to just the SW mode, we avoid the issue and can
> > > allow recursive printing?
> >
> > Thanks for the suggestion! But based on the suggestion, I tried to
> > change code as below, the console even cannot work when boot the
> > kernel:
> >
> > static void msm_reset_dm_count(struct uart_port *port, int count)
> > {
> > + u32 val;
> > +
> > msm_wait_for_xmitr(port);
> > - msm_write(port, count, UARTDM_NCF_TX);
> > - msm_read(port, UARTDM_NCF_TX);
> > +
> > + val = msm_read(port, UARTDM_DMEN);
> > +
> > + /*
> > + * NCF is only enabled for SW transmit mode and is
> > + * skipped for BAM mode.
> > + */
> > + if (!(val & UARTDM_DMEN_TX_BAM_ENABLE) &&
> > + !(val & UARTDM_DMEN_RX_BAM_ENABLE)) {
> > + msm_write(port, count, UARTDM_NCF_TX);
> > + msm_read(port, UARTDM_NCF_TX);
> > + }
> > }
> >
> >
> > Alternatively, when exit from __msm_console_write() and if detect the
> > case for without acquiring spinlock, invoke msm_wait_for_xmitr() to wait
> > for transmit completion looks a good candidate solution. The updated
> > patch is as below. Please let me know if this is doable?
> >
> > diff --git a/drivers/tty/serial/msm_serial.c b/drivers/tty/serial/msm_serial.c
> > index 1db79ee8a886..aa6a494c898d 100644
> > --- a/drivers/tty/serial/msm_serial.c
> > +++ b/drivers/tty/serial/msm_serial.c
> > @@ -190,6 +190,7 @@ struct msm_port {
> > bool break_detected;
> > struct msm_dma tx_dma;
> > struct msm_dma rx_dma;
> > + struct cpumask curr_user;
> > };
> >
> > #define UART_TO_MSM(uart_port) container_of(uart_port, struct msm_port, uart)
> > @@ -440,6 +441,7 @@ static void msm_complete_tx_dma(void *args)
> > u32 val;
> >
> > spin_lock_irqsave(&port->lock, flags);
> > + cpumask_set_cpu(smp_processor_id(), &msm_port->curr_user);
> >
> > /* Already stopped */
> > if (!dma->count)
> > @@ -474,6 +476,7 @@ static void msm_complete_tx_dma(void *args)
> >
> > msm_handle_tx(port);
> > done:
> > + cpumask_clear_cpu(smp_processor_id(), &msm_port->curr_user);
> > spin_unlock_irqrestore(&port->lock, flags);
> > }
> >
> > @@ -548,6 +551,7 @@ static void msm_complete_rx_dma(void *args)
> > u32 val;
> >
> > spin_lock_irqsave(&port->lock, flags);
> > + cpumask_set_cpu(smp_processor_id(), &msm_port->curr_user);
> >
> > /* Already stopped */
> > if (!dma->count)
> > @@ -594,6 +598,7 @@ static void msm_complete_rx_dma(void *args)
> >
> > msm_start_rx_dma(msm_port);
> > done:
> > + cpumask_clear_cpu(smp_processor_id(), &msm_port->curr_user);
> > spin_unlock_irqrestore(&port->lock, flags);
> >
> > if (count)
> > @@ -932,6 +937,7 @@ static irqreturn_t msm_uart_irq(int irq, void *dev_id)
> > u32 val;
> >
> > spin_lock_irqsave(&port->lock, flags);
> > + cpumask_set_cpu(smp_processor_id(), &msm_port->curr_user);
> > misr = msm_read(port, UART_MISR);
> > msm_write(port, 0, UART_IMR); /* disable interrupt */
> >
> > @@ -963,6 +969,7 @@ static irqreturn_t msm_uart_irq(int irq, void *dev_id)
> > msm_handle_delta_cts(port);
> >
> > msm_write(port, msm_port->imr, UART_IMR); /* restore interrupt */
> > + cpumask_clear_cpu(smp_processor_id(), &msm_port->curr_user);
> > spin_unlock_irqrestore(&port->lock, flags);
> >
> > return IRQ_HANDLED;
> > @@ -1573,10 +1580,12 @@ static inline struct uart_port *msm_get_port_from_line(unsigned int line)
> > static void __msm_console_write(struct uart_port *port, const char *s,
> > unsigned int count, bool is_uartdm)
> > {
> > + struct msm_port *msm_port = UART_TO_MSM(port);
> > int i;
> > int num_newlines = 0;
> > bool replaced = false;
> > void __iomem *tf;
> > + int locked = 1;
> >
> > if (is_uartdm)
> > tf = port->membase + UARTDM_TF;
> > @@ -1589,7 +1598,15 @@ static void __msm_console_write(struct uart_port *port, const char *s,
> > num_newlines++;
> > count += num_newlines;
> >
> > - spin_lock(&port->lock);
> > + if (port->sysrq)
> > + locked = 0;
> > + else if (oops_in_progress)
> > + locked = spin_trylock(&port->lock);
> > + else if (cpumask_test_cpu(smp_processor_id(), &msm_port->curr_user))
> > + locked = 0;
> > + else
> > + spin_lock(&port->lock);
> > +
> > if (is_uartdm)
> > msm_reset_dm_count(port, count);
> >
> > @@ -1625,7 +1642,12 @@ static void __msm_console_write(struct uart_port *port, const char *s,
> > iowrite32_rep(tf, buf, 1);
> > i += num_chars;
> > }
> > - spin_unlock(&port->lock);
> > +
> > + if (!locked)
> > + msm_wait_for_xmitr(port);
>
> Sorry, catching up from some travel.
>
> I don't understand this. At this point, haven't we already called
> msm_reset_dm_count() and "corrupted" the state of the hardware?

Yeah, at here msm_reset_dm_count() has been called.

msm_wait_for_xmitr() is used to wait for completing transmition.
So we can get flow as:

msm_complete_tx_dma()
kmalloc() fail
__msm_console_write()
msm_reset_dm_count()
output logs
msm_wait_for_xmitr() => ensure to not impact out flow

My essential reason for adding msm_wait_for_xmitr() is to cleanup
the "corrupted" state before return to out flow.

Thanks,
Leo Yan

> > +
> > + if (locked)
> > + spin_unlock(&port->lock);
> > }