> -----Original Message-----
> From: Dexuan Cui
> Sent: Sunday, September 6, 2015 4:48 AM
> To: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>; Vitaly Kuznetsov
> <vkuznets@xxxxxxxxxx>
> Cc: Jiri Slaby <jslaby@xxxxxxxx>; linux-serial@xxxxxxxxxxxxxxx; linux-
> kernel@xxxxxxxxxxxxxxx; KY Srinivasan <kys@xxxxxxxxxxxxx>; Peter Hurley
> <peter@xxxxxxxxxxxxxxxxxx>
> Subject: RE: [PATCH] serial: core: prevent softlockups on slow consoles
>
> > -----Original Message-----
> > From: Greg Kroah-Hartman
> > Sent: Saturday, September 5, 2015 0:10
> > On Fri, Sep 04, 2015 at 09:19:38AM +0200, Vitaly Kuznetsov wrote:
> > > Greg Kroah-Hartman writes:
> > >
> > > > On Mon, Aug 31, 2015 at 04:34:16PM +0200, Vitaly Kuznetsov wrote:
> > > >> Hyper-V serial port is very slow on multi-vCPU guest, this causes
> > > >> soflockups on intensive console writes. Touch nmi watchdog after
> putting
> > > >> every char on port to avoid the issue for all serial drivers, the overhead
> > > >> should be small.
> > > >>
> > > >> This is just a part of the fix: serial8250_console_write() disables irqs
> > > >> for all its execution time (which on such slow consoles can be dozens
> of
> > > >> seconds), it should be possible to observe devices being stuck on this
> > > >> CPU. We need to find a better way, e.g. do output in batches enabling
> irqs
> > > >> in between.
> > > >>
> > > >> Signed-off-by: Vitaly Kuznetsov
>
> Thank you Vitaly for the help of trying to mitigate the issue!
>
> Please let me explain the "real" issue here since I investigated the same issue
> a
> few months ago.
>
> (Please see the below)
>
> > > >> ---
> > > >> drivers/tty/serial/serial_core.c | 3 ++-
> > > >> 1 file changed, 2 insertions(+), 1 deletion(-)
> > > >>
> > > >> diff --git a/drivers/tty/serial/serial_core.c
> b/drivers/tty/serial/serial_core.c
> > > >> index f368520..cc05785 100644
> > > >> --- a/drivers/tty/serial/serial_core.c
> > > >> +++ b/drivers/tty/serial/serial_core.c
> > > >> @@ -33,7 +33,7 @@
> > > >> #include <linux/serial.h> /* for serial_state and serial_icounter_struct
> */
> > > >> #include <linux/serial_core.h>
> > > >> #include <linux/delay.h>
> > > >> -#include <linux/mutex.h>
> > > >> +#include <linux/nmi.h>
> > > >>
> > > >> #include <asm/irq.h>
> > > >> #include <asm/uaccess.h>
> > > >> @@ -1792,6 +1792,7 @@ void uart_console_write(struct uart_port
> *port,
> > const char *s,
> > > >> if (*s == '\n')
> > > >> putchar(port, '\r');
> > > >> putchar(port, *s);
> > > >> + touch_nmi_watchdog();
> > > >
> > > > I don't like this, please narrow this down to the real problem that your
> > > > hardware has here, the putchar function should not be this slow. If it
> > > > is, something is wrong.
> > >
> > > I'm afraid this is really the case:
> > >
> > > 3) | serial8250_console_putchar() {
> > > 3) | wait_for_xmitr() {
> > > 3) # 3111.189 us | io_serial_in();
> > > 3) # 3115.334 us | }
> > > 3) # 2234.099 us | io_serial_out();
> > > 3) # 5353.883 us | }
> > >
> > > This is one char and I use local pipe for Hyper-V output. In case
> > > something like remote pipe is in use ...
> > >
> > > So I'm sorry, but I don't really understand the suggestion to 'narrow
> > > this down' - this is how slow Hyper-V serial's implementation is,
> > > io_serial_in() is just an inb() and io_serial_out() is an outb().
> >
> > So a call to inb() and outb() really takes that long? Again, this is
>
> Yes, if you're using a VM with many vCPUs, like 16 or 32 vCPUs.
> If you only use 1 vCPU, inb()/outb() is pretty fast as it should be.
> The more vCPU your VM has, the slower inb()/outb() can be.
> There is almost a linear relationship here...
>
> > broken somewhere in the hypervisor, or you need to fix up the platform
> Yes, the serial emulation code in the host is broken for SMP guest.
>
> Historically, usually Windows VM itself doesn't use the serial so much
> as Linux VM. The most important usage of the serial in Windows VM is
> windbg: a host debugger can connect to the VM by its (virtual) serial.
>
> Windbg may use multiple consecutive ins/outs instructions, trying to
> exchange data faster between the host and Windows VM. In the host's
> serial emulation code, there is a software instruction emulator, which
> tries to "execute" the VM's ins/outs on behalf of the VM -- this way,
> there are fewer ins/outs intercepts to the hypervisor (in Intel CPU, it's
> called "VM exit") and the intercepts are forwarded to the host's serial
> emulation code.
>
> This optimization of reducing the number of the intercepts is probably
> good for the 6-years-ago old CPUs, but is pretty questionable for today's
> CPUs since the cost of the intercept has been reduced really a lot.
>
> A side effect of the software instruction emulator in the host's serial
> emulation code is: it triggers the need to pause the other vCPUs when
> emulating ins/outs, probably for the atomicity of accessing the
> memory(?). Unluckily it turns out pausing n vCPUs is expensive,
> especially when n is >8 and on relatively new faster CPUs.
> I suspect nobody ever tested the case of "vCPUS > 8" here.
>
> This is the cause of the slow serial issue here, AFAIK.
>
> > logic for inb() and outb() to properly kick the watchdog. Perhaps
> > hyperv needs its own arch type for this kind of crud?
> >
> > Don't "paper over" the real issue here please.
> >
> > greg k-h
>
> I agree with Greg.
>
> AFAIK, the "slow serial console for SMP guest" issue should be fixed
> in Hyper-V 2016. Unluckily IMO there is no workaround for the
> current version of Hyper-V -- we'd better avoid outputting lots of
> messages by the serial console in a SMP Hyper-V VM with many vCPUs.
The fix is in Server 2016 (to address the needs of Linux). We are looking at potentially
backporting the host side fix.
K. Y
>
> Thanks,
> -- Dexuan
--
To unsubscribe from this list: send the line "unsubscribe linux-serial" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
