| Message ID | 20220401194645.1738747-2-wander@redhat.com |
|---|---|
| State | Superseded |
| Headers | show |
| Series | [v6] serial/8250: Use fifo in 8250 console driver | expand |
On Fri, Apr 01, 2022 at 04:46:42PM -0300, Wander Lairson Costa wrote: > Note: I am using a small test app + driver located at [0] for the > problem description. serco is a driver whose write function dispatches > to the serial controller. sertest is a user-mode app that writes n bytes > to the serial console using the serco driver. > > While investigating a bug in the RHEL kernel, I noticed that the serial > console throughput is way below the configured speed of 115200 bps in > a HP Proliant DL380 Gen9. I was expecting something above 10KB/s, but > I got 2.5KB/s. > > $ time ./sertest -n 2500 /tmp/serco > > real 0m0.997s > user 0m0.000s > sys 0m0.997s > > With the help of the function tracer, I then noticed the serial > controller was taking around 410us seconds to dispatch one single byte: > > $ trace-cmd record -p function_graph -g serial8250_console_write \ > ./sertest -n 1 /tmp/serco > > $ trace-cmd report > > | serial8250_console_write() { > 0.384 us | _raw_spin_lock_irqsave(); > 1.836 us | io_serial_in(); > 1.667 us | io_serial_out(); > | uart_console_write() { > | serial8250_console_putchar() { > | wait_for_xmitr() { > 1.870 us | io_serial_in(); > 2.238 us | } > 1.737 us | io_serial_out(); > 4.318 us | } > 4.675 us | } > | wait_for_xmitr() { > 1.635 us | io_serial_in(); > | __const_udelay() { > 1.125 us | delay_tsc(); > 1.429 us | } > ... > ... > ... > 1.683 us | io_serial_in(); > | __const_udelay() { > 1.248 us | delay_tsc(); > 1.486 us | } > 1.671 us | io_serial_in(); > 411.342 us | } > > In another machine, I measured a throughput of 11.5KB/s, with the serial > controller taking between 80-90us to send each byte. That matches the > expected throughput for a configuration of 115200 bps. > > This patch changes the serial8250_console_write to use the 16550 fifo > if available. In my benchmarks I got around 25% improvement in the slow > machine, and no performance penalty in the fast machine. ... > + use_fifo = (up->capabilities & UART_CAP_FIFO) && > + /* > + * BCM283x requires to check the fifo > + * after each byte. > + */ > + !(up->capabilities & UART_CAP_MINI) && Perhaps you need to also comment why we are using tx_loadsz and not fifosize. > + up->tx_loadsz > 1 && > + (up->fcr & UART_FCR_ENABLE_FIFO) && > + port->state && > + test_bit(TTY_PORT_INITIALIZED, &port->state->port.iflags) && > + /* > + * After we put a data in the fifo, the controller will send > + * it regardless of the CTS state. Therefore, only use fifo > + * if we don't use control flow. > + */ > + !(up->port.flags & UPF_CONS_FLOW);
On Mon, Apr 4, 2022 at 6:32 AM Andy Shevchenko <andriy.shevchenko@linux.intel.com> wrote: > > On Fri, Apr 01, 2022 at 04:46:42PM -0300, Wander Lairson Costa wrote: > > Note: I am using a small test app + driver located at [0] for the > > problem description. serco is a driver whose write function dispatches > > to the serial controller. sertest is a user-mode app that writes n bytes > > to the serial console using the serco driver. > > > > While investigating a bug in the RHEL kernel, I noticed that the serial > > console throughput is way below the configured speed of 115200 bps in > > a HP Proliant DL380 Gen9. I was expecting something above 10KB/s, but > > I got 2.5KB/s. > > > > $ time ./sertest -n 2500 /tmp/serco > > > > real 0m0.997s > > user 0m0.000s > > sys 0m0.997s > > > > With the help of the function tracer, I then noticed the serial > > controller was taking around 410us seconds to dispatch one single byte: > > > > $ trace-cmd record -p function_graph -g serial8250_console_write \ > > ./sertest -n 1 /tmp/serco > > > > $ trace-cmd report > > > > | serial8250_console_write() { > > 0.384 us | _raw_spin_lock_irqsave(); > > 1.836 us | io_serial_in(); > > 1.667 us | io_serial_out(); > > | uart_console_write() { > > | serial8250_console_putchar() { > > | wait_for_xmitr() { > > 1.870 us | io_serial_in(); > > 2.238 us | } > > 1.737 us | io_serial_out(); > > 4.318 us | } > > 4.675 us | } > > | wait_for_xmitr() { > > 1.635 us | io_serial_in(); > > | __const_udelay() { > > 1.125 us | delay_tsc(); > > 1.429 us | } > > ... > > ... > > ... > > 1.683 us | io_serial_in(); > > | __const_udelay() { > > 1.248 us | delay_tsc(); > > 1.486 us | } > > 1.671 us | io_serial_in(); > > 411.342 us | } > > > > In another machine, I measured a throughput of 11.5KB/s, with the serial > > controller taking between 80-90us to send each byte. That matches the > > expected throughput for a configuration of 115200 bps. > > > > This patch changes the serial8250_console_write to use the 16550 fifo > > if available. In my benchmarks I got around 25% improvement in the slow > > machine, and no performance penalty in the fast machine. > > ... > > > + use_fifo = (up->capabilities & UART_CAP_FIFO) && > > + /* > > + * BCM283x requires to check the fifo > > + * after each byte. > > + */ > > + !(up->capabilities & UART_CAP_MINI) && > > Perhaps you need to also comment why we are using tx_loadsz and not fifosize. > Maybe it is better to document their difference in the struct declaration and not in a random usage. [snip]
On Mon, Apr 04, 2022 at 10:27:30AM -0300, Wander Costa wrote: > On Mon, Apr 4, 2022 at 6:32 AM Andy Shevchenko > <andriy.shevchenko@linux.intel.com> wrote: > > On Fri, Apr 01, 2022 at 04:46:42PM -0300, Wander Lairson Costa wrote: ... > > > + use_fifo = (up->capabilities & UART_CAP_FIFO) && > > > + /* > > > + * BCM283x requires to check the fifo > > > + * after each byte. > > > + */ > > > + !(up->capabilities & UART_CAP_MINI) && > > > > Perhaps you need to also comment why we are using tx_loadsz and not fifosize. > > Maybe it is better to document their difference in the struct > declaration and not in a random usage. Here, when one reads a code (as a non-familiar with the area), the use of tx_loadsz confuses if one saw previously fifosize used somewhere. So, I agree that it's good to document in the structure, but here it's also good to have a comment to briefly hint the reader why this and not the other one is used.
diff --git a/drivers/tty/serial/8250/8250_port.c b/drivers/tty/serial/8250/8250_port.c index 318af6f13605..8113e6c73407 100644 --- a/drivers/tty/serial/8250/8250_port.c +++ b/drivers/tty/serial/8250/8250_port.c @@ -2077,10 +2077,7 @@ static void serial8250_break_ctl(struct uart_port *port, int break_state) serial8250_rpm_put(up); } -/* - * Wait for transmitter & holding register to empty - */ -static void wait_for_xmitr(struct uart_8250_port *up, int bits) +static void wait_for_lsr(struct uart_8250_port *up, int bits) { unsigned int status, tmout = 10000; @@ -2097,6 +2094,16 @@ static void wait_for_xmitr(struct uart_8250_port *up, int bits) udelay(1); touch_nmi_watchdog(); } +} + +/* + * Wait for transmitter & holding register to empty + */ +static void wait_for_xmitr(struct uart_8250_port *up, int bits) +{ + unsigned int tmout; + + wait_for_lsr(up, bits); /* Wait up to 1s for flow control if necessary */ if (up->port.flags & UPF_CONS_FLOW) { @@ -3332,6 +3339,35 @@ static void serial8250_console_restore(struct uart_8250_port *up) serial8250_out_MCR(up, UART_MCR_DTR | UART_MCR_RTS); } +/* + * Print a string to the serial port using the device FIFO + * + * It sends fifosize bytes and then waits for the fifo + * to get empty. + */ +static void serial8250_console_fifo_write(struct uart_8250_port *up, + const char *s, unsigned int count) +{ + int i; + const char *end = s + count; + unsigned int fifosize = up->tx_loadsz; + bool cr_sent = false; + + while (s != end) { + wait_for_lsr(up, UART_LSR_THRE); + + for (i = 0; i < fifosize && s != end; ++i) { + if (*s == '\n' && !cr_sent) { + serial_out(up, UART_TX, '\r'); + cr_sent = true; + } else { + serial_out(up, UART_TX, *s++); + cr_sent = false; + } + } + } +} + /* * Print a string to the serial port trying not to disturb * any possible real use of the port... @@ -3347,7 +3383,7 @@ void serial8250_console_write(struct uart_8250_port *up, const char *s, struct uart_8250_em485 *em485 = up->em485; struct uart_port *port = &up->port; unsigned long flags; - unsigned int ier; + unsigned int ier, use_fifo; int locked = 1; touch_nmi_watchdog(); @@ -3379,7 +3415,27 @@ void serial8250_console_write(struct uart_8250_port *up, const char *s, mdelay(port->rs485.delay_rts_before_send); } - uart_console_write(port, s, count, serial8250_console_putchar); + use_fifo = (up->capabilities & UART_CAP_FIFO) && + /* + * BCM283x requires to check the fifo + * after each byte. + */ + !(up->capabilities & UART_CAP_MINI) && + up->tx_loadsz > 1 && + (up->fcr & UART_FCR_ENABLE_FIFO) && + port->state && + test_bit(TTY_PORT_INITIALIZED, &port->state->port.iflags) && + /* + * After we put a data in the fifo, the controller will send + * it regardless of the CTS state. Therefore, only use fifo + * if we don't use control flow. + */ + !(up->port.flags & UPF_CONS_FLOW); + + if (likely(use_fifo)) + serial8250_console_fifo_write(up, s, count); + else + uart_console_write(port, s, count, serial8250_console_putchar); /* * Finally, wait for transmitter to become empty
Note: I am using a small test app + driver located at [0] for the problem description. serco is a driver whose write function dispatches to the serial controller. sertest is a user-mode app that writes n bytes to the serial console using the serco driver. While investigating a bug in the RHEL kernel, I noticed that the serial console throughput is way below the configured speed of 115200 bps in a HP Proliant DL380 Gen9. I was expecting something above 10KB/s, but I got 2.5KB/s. $ time ./sertest -n 2500 /tmp/serco real 0m0.997s user 0m0.000s sys 0m0.997s With the help of the function tracer, I then noticed the serial controller was taking around 410us seconds to dispatch one single byte: $ trace-cmd record -p function_graph -g serial8250_console_write \ ./sertest -n 1 /tmp/serco $ trace-cmd report | serial8250_console_write() { 0.384 us | _raw_spin_lock_irqsave(); 1.836 us | io_serial_in(); 1.667 us | io_serial_out(); | uart_console_write() { | serial8250_console_putchar() { | wait_for_xmitr() { 1.870 us | io_serial_in(); 2.238 us | } 1.737 us | io_serial_out(); 4.318 us | } 4.675 us | } | wait_for_xmitr() { 1.635 us | io_serial_in(); | __const_udelay() { 1.125 us | delay_tsc(); 1.429 us | } ... ... ... 1.683 us | io_serial_in(); | __const_udelay() { 1.248 us | delay_tsc(); 1.486 us | } 1.671 us | io_serial_in(); 411.342 us | } In another machine, I measured a throughput of 11.5KB/s, with the serial controller taking between 80-90us to send each byte. That matches the expected throughput for a configuration of 115200 bps. This patch changes the serial8250_console_write to use the 16550 fifo if available. In my benchmarks I got around 25% improvement in the slow machine, and no performance penalty in the fast machine. Signed-off-by: Wander Lairson Costa <wander@redhat.com> --- drivers/tty/serial/8250/8250_port.c | 68 ++++++++++++++++++++++++++--- 1 file changed, 62 insertions(+), 6 deletions(-)