Serial console hangs with Linux 2.6.20 HVM guest

Serial console hangs with Linux 2.6.20 HVM guest

[Anders Kaseorg]

I am seeing a problem with the Xen emulated serial console.  When
running a Linux 2.6.20 HVM guest that has CONFIG_HOTPLUG_CPU=n, the
guest blocks on output to the console until it receives input keypresses
from `xm console`.  This prevents the guest from booting up without
banging on some keys, and makes interactive use of the console
difficult.

By bisecting Linux kernel commits, I found that the bug goes away in
commit 40b36daad0ac704e6d5c1b75789f371ef5b053c1 (v2.6.21-rc1~261), which
is a workaround for buggy UARTs on certain HP machines.

http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=40b36daad0ac704e6d5c1b75789f371ef5b053c1

“The patch below works around a minor bug found in the UART of the
remote management card used in many HP ia64 and parisc servers (aka the
Diva UARTs).  The problem is that the UART does not reassert the THRE
interrupt if it has been previously cleared and the IIR THRI bit is
re-enabled.  This can produce a very annoying failure mode when used as
a serial console, allowing a boot/reboot to hang indefinitely until an
RX interrupt kicks it into working again (ie. an unattended reboot could
stall).”

That matches my symptoms exactly, which suggests that the Xen UART
probably has a similar bug.

I’ve seen this in Xen 3.2.1 and 3.3.1.  (My host is running Debian Lenny
amd64, with the Xen dom0 kernel 2.6.24-23-xen from Ubuntu Hardy, on a
server with two quad-core Xeons.)

Anders

Re: Serial console hangs with Linux 2.6.20 HVM guest

[Ian Jackson]

Anders Kaseorg writes ("[Xen-devel] Serial console hangs with Linux 2.6.20 HVM guest"):
> http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=40b36daad0ac704e6d5c1b75789f371ef5b053c1

Thanks for this.  However looking at the code I'm not sure that's
what's actually going on.  Very old versions of qemu did have the bug
that patch describes, but it was fixed in qemu upstream in 2004 in
this commit:

  commit 60e336dbb837ef4d5053433f9ee391feb102be36
  Author: bellard <bellard@c046a42c-6fe2-441c-8c8c-71466251a162>
  Date:   Tue Aug 24 21:55:28 2004 +0000

      serial interrupt fix (Hampa Hug)

      git-svn-id: svn://svn.savannah.nongnu.org/qemu/trunk@1049 c046a42c-6fe2-441c-8c8c-71466251a162

As far as I can see this fix is included in all relevant versions of
qemu.  The upstream Linux changeset seems like it would fix any
general IRQ loss problem with the UART.

If you can reproduce this problem, I would like to try to debug it.

Probably the best way is to attach to the running qemu with gdb:
  * Add
       CFLAGS += -g -O0
    to xen-unstable.hg/.config (creating it if necessary) and rebuild
    and install the resulting qemu-dm binary where it will be built.
  * Start the VM in the usual way.
  * cd to the qemu-xen-unstable tree (probably tools/ioemu-remote);
    use ps to find the pid of qemu-dm, and say
        gdb i386-dm/qemu-dm <pid>
    and then at the gdb prompt say
        handle SIGUSR2 nostop noprint
        break serial_ioport_write if (addr&7)==1
        cont
  * do whatever it is that makes the VM stuck
  * when it next stops it will be in serial_ioport_write setting
    the IER.  So
        print val
        print *s
    and email the output to this list.

NB I haven't tested this recipe.  If you get syntax errors or
something doesn't work I'll actually do so and let you know what to
type.

Ian.

Re: Serial console hangs with Linux 2.6.20 HVM guest

[Anders Kaseorg]

On Thu, 2009-02-05 at 17:04 +0000, Ian Jackson wrote:
>         handle SIGUSR2 nostop noprint
>         break serial_ioport_write if (addr&7)==1
>         cont
>   * do whatever it is that makes the VM stuck
>   * when it next stops it will be in serial_ioport_write setting
>     the IER.  So
>         print val
>         print *s

This breakpoint is triggered for all messages printed by the kernel,
which always showed up with no delay; but it is only occasionally
triggered for strings printed by userspace, even after forcing those
strings to show up by sending keystrokes.

Here is one of the latter cases.  (I am sitting at a
“root@andersk-intrepid:~# ” prompt, repeatedly pressing Enter.  Each
keypress causes the previous prompt to show up, followed by a newline,
and the current prompt is stalled.)

Breakpoint 1, serial_ioport_write (opaque=0xb342e0, addr=1, val=5)
    at /home/andersk/xen-3-3.3.1/debian/build/build-utils_amd64/tools/ioemu-dir/hw/serial.c:413
413	{
(gdb) print val
$5 = 5
(gdb) print *s
$6 = {divider = 1, rbr = 0 '\0', thr = 32 ' ', tsr = 32 ' ', ier = 5 '\005', iir = 193 '�', 
  lcr = 19 '\023', mcr = 11 '\v', lsr = 96 '`', msr = 176 '�', scr = 0 '\0', fcr = 129 '\201', 
  thr_ipending = 1, irq = 0xb1d610, chr = 0xb122a0, last_break_enable = 0, base = 0, 
  it_shift = 0, baudbase = 115200, tsr_retry = 0, last_xmit_ts = 380482341502, recv_fifo = {
    data = '\r' <repeats 16 times>, count = 0 '\0', itl = 8 '\b', tail = 0 '\0', 
    head = 0 '\0'}, xmit_fifo = {data = "repid:~# rsk-int", count = 0 '\0', itl = 0 '\0', 
    tail = 9 '\t', head = 9 '\t'}, fifo_timeout_timer = 0xb31ad0, timeout_ipending = 0, 
  transmit_timer = 0xb31b00, char_transmit_time = 78120, poll_msl = -1, 
  modem_status_poll = 0xb327e0}

Anders

Re: Serial console hangs with Linux 2.6.20 HVM guest

[Anders Kaseorg]

In case this is more interesting, here is a log of all the breakpoints
hit during a Linux boot, up to the point where it hangs:

http://web.mit.edu/andersk/Public/xen-serial-log

Anders

Re: Serial console hangs with Linux 2.6.20 HVM guest

[Ian Jackson]

Anders Kaseorg writes ("Re: [Xen-devel] Serial console hangs with Linux 2.6.20 HVM guest"):
> This breakpoint is triggered for all messages printed by the kernel,
> which always showed up with no delay; but it is only occasionally
> triggered for strings printed by userspace, even after forcing those
> strings to show up by sending keystrokes.

That's interesting ....

Anders Kaseorg writes ("Re: [Xen-devel] Serial console hangs with Linux 2.6.20 HVM guest"):
> In case this is more interesting, here is a log of all the breakpoints
> hit during a Linux boot, up to the point where it hangs:
> 
> http://web.mit.edu/andersk/Public/xen-serial-log

.... looking at this, it seems like we definitely don't have the bug
that the Linux kernel is trying to work around.  In each of the cases
where the IER was written there by the kernel, an interrupt was
already pending.

But since the Linux change is just to have a failsafe timer for lost
interrupts, any other kind of lost interrupt situation might cause it.
Perhaps the kernel is expecting us to deassert and reassert the
interrupt in some situation when we aren't; since the interrupts are
converted from level to edge triggered it's possible that we are
failing to _deassert_ the interrupt when we should.  I'm just
speculating here at the moment ...

Can you recompile your qemu-dm with
  #define DEBUG_SERIAL
near the top of serial.c uncommented, and try booting up and get it to
the point where some output is supposed to have come out but has got
stuck, and then send me your /var/log/xen/qemu-dm-<whatever>.log ?

If it's short post it to the list; otherwise put it up on a webpage or
email it to me privately.

Thanks,
Ian.

Re: Serial console hangs with Linux 2.6.20 HVM guest

[Anders Kaseorg]

On Mon, 2009-02-09 at 17:57 +0000, Ian Jackson wrote:
> Can you recompile your qemu-dm with
>   #define DEBUG_SERIAL
> near the top of serial.c uncommented, and try booting up and get it to
> the point where some output is supposed to have come out but has got
> stuck, and then send me your /var/log/xen/qemu-dm-<whatever>.log ?

Sure: http://web.mit.edu/andersk/Public/qemu-dm-andersk-intrepid.log (3
MB).

Anders

Re: Serial console hangs with Linux 2.6.20 HVM guest

[Ian Jackson]

Anders Kaseorg writes ("Re: [Xen-devel] Serial console hangs with Linux 2.6.20 HVM guest"):
> In case this is more interesting, here is a log of all the breakpoints
> hit during a Linux boot, up to the point where it hangs:
> 
> http://web.mit.edu/andersk/Public/xen-serial-log

Thanks, that's great.  But I think it's a kernel bug.  The last thing
the kernel does is read the IIR (Interrupt Identification Register)
twice at a time when the transmit FIFO is empty.  Reading the IIR is
(sadly) not a side-effect-free operation; specifically, it cancels any
outstanding transmit fifo/buffer empty interrupt[1].  So the first
time it gets told `Transmit Holding Register Empty interrupt', but
that has the effect of clearing the interrupt so the second time it
reads the IIRC it gets `no interrupt pending'.

[1] I'm getting this out of the National Semiconductor datasheet for
the PC16550D, document number TL/C/8652, June 1995.  See for example
section 8.11 `FIFO Interrupt Mode Operation' item A:
   The transmit holding register interrupt (02) occurs when
   the XMIT FIFO is empty; it is cleared as soon as the transmitter
   holding register is written to ([...]) or the IIR is read.
As far as I can see qemu-dm is emulating this accurately.

Can you point me at the exact kernel source code you're using ?

Thanks,
Ian.

Re: Serial console hangs with Linux 2.6.20 HVM guest

[Anders Kaseorg]

On Tue, 10 Feb 2009, Ian Jackson wrote:
> Can you point me at the exact kernel source code you're using ?

Yes, I took Linux v2.6.20 on amd64, ran `make defconfig`, then ran `make 
menuconfig` and turned off CONFIG_HOTPLUG_CPU (Processor type and features 
→ Support for hot-pluggable CPUs).

Anders

[PATCH] IRQ handling race and spurious IIR read in serial/8250.c

[Ian Jackson]

Anders Kaseorg writes ("Re: Serial console hangs with Linux 2.6.20 HVM guest"):
> Yes, I took Linux v2.6.20 on amd64, ran `make defconfig`, then ran `make 
> menuconfig` and turned off CONFIG_HOTPLUG_CPU (Processor type and features 
> ò Support for hot-pluggable CPUs).

Thanks.  I think I have tracked down the bug.  In
drivers/serial/8250.c in Linux there are two bugs:
 1. UART_BUG_TXEN can be spuriously set, due to an IRQ race
 2. The workaround then applied by the kernel is itself buggy

Anders: can you try two tests for me ?  Firstly, in
serial8250_startup, delete the section which sets UART_BUG_TXEN (see
2nd patch below); I think this will fix the symptoms for you.
Secondly, in serial8250_start_tx delete the read from the IIR and the
relevant branch of the text (see 3rd patch below); I think this will
also in itself fix your symptoms.  I haven't compiled either patch (so
you may find that eg I missed deleting some variables).


The bugs in detail (this discussion applies to 2.6.20 and also to
2.6.28.4):

 1. The hunk of serial8250_startup I quote below attempts to discover
    whether writing the IER re-asserts the THRI (transmit ready)
    interrupt.  However the spinlock that it has taken out,
    port->lock, is not the one that the IRQ service routine takes
    before reading the IIR (i->lock).  As a result, on an SMP system
    the generated interrupt races with the straight-line code in
    serial8250_startup.

    If serial8250_startup loses the race (perhaps because the system
    is a VM and its VCPU got preempted), UART_BUG_TXEN is spuriously
    added to bugs.  This is quite unlikely in a normal system but in
    certain Xen configurations, particularly ones where there is CPU
    pressure, we may lose the race every time.

    It is not exactly clear to me how this ought to be resolved.  One
    possibility is that the UART_BUG_TXEN problem might be worked
    around perfectly well by the new and very similar workaround
    UART_BUG_THRE[1] in 2.6.21ish in which case it could just be
    removed.
    
 2. UART_BUG_TXEN's workaround appears to be intended to be harmless.
    However what it actually does is to read the IIR, thus clearing
    any actual interrupt (including incidentally non-THRI), and then
    only perform the intended servicing if the interrupt was _not_
    asserted.  That is, it breaks on any serial port with the bug.

    As far as I can see there is not much use in UART_BUG_TXEN reading
    IIR at all, so a suitable change if we want to keep UART_BUG_TXEN
    might be the first patch I enclose below (again, not compiled
    or tested).

    If UART_BUG_TXEN is retained something along these lines should be
    done at the very least.

Ian.

[1] http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=40b36daad0ac704e6d5c1b75789f371ef5b053c1
    in which case UART


Proposed initial band-aid fix (against 2.6.28.4):


Do not read IIR in serial8250_start_tx when UART_BUG_TXEN

Reading the IIR clears some oustanding interrupts so it is not safe.
Instead, simply transmit immediately if the buffer is empty without
regard to IIR.

Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>

--- ../linux-2.6.28.4/drivers/serial/8250.c~	2009-02-06 21:47:45.000000000 +0000
+++ ../linux-2.6.28.4/drivers/serial/8250.c	2009-02-11 15:55:24.000000000 +0000
@@ -1257,14 +1257,12 @@
 		serial_out(up, UART_IER, up->ier);
 
 		if (up->bugs & UART_BUG_TXEN) {
-			unsigned char lsr, iir;
+			unsigned char lsr;
 			lsr = serial_in(up, UART_LSR);
 			up->lsr_saved_flags |= lsr & LSR_SAVE_FLAGS;
-			iir = serial_in(up, UART_IIR) & 0x0f;
 			if ((up->port.type == PORT_RM9000) ?
-				(lsr & UART_LSR_THRE &&
-				(iir == UART_IIR_NO_INT || iir == UART_IIR_THRI)) :
-				(lsr & UART_LSR_TEMT && iir & UART_IIR_NO_INT))
+				(lsr & UART_LSR_THRE) :
+				(lsr & UART_LSR_TEMT))
 				transmit_chars(up);
 		}
 	}


Anders - first patch to try (against 2.6.20):

--- drivers/serial/8250.c~	2007-02-04 18:44:54.000000000 +0000
+++ drivers/serial/8250.c	2009-02-11 15:39:43.000000000 +0000
@@ -1645,25 +1645,6 @@
 
 	serial8250_set_mctrl(&up->port, up->port.mctrl);
 
-	/*
-	 * Do a quick test to see if we receive an
-	 * interrupt when we enable the TX irq.
-	 */
-	serial_outp(up, UART_IER, UART_IER_THRI);
-	lsr = serial_in(up, UART_LSR);
-	iir = serial_in(up, UART_IIR);
-	serial_outp(up, UART_IER, 0);
-
-	if (lsr & UART_LSR_TEMT && iir & UART_IIR_NO_INT) {
-		if (!(up->bugs & UART_BUG_TXEN)) {
-			up->bugs |= UART_BUG_TXEN;
-			pr_debug("ttyS%d - enabling bad tx status workarounds\n",
-				 port->line);
-		}
-	} else {
-		up->bugs &= ~UART_BUG_TXEN;
-	}
-
 	spin_unlock_irqrestore(&up->port.lock, flags);
 
 	/*


Anders - second patch to try (against 2.6.20):
Fix should be suitable for distribution IMO.

Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>

--- drivers/serial/8250.c~	2007-02-04 18:44:54.000000000 +0000
+++ drivers/serial/8250.c	2009-02-11 15:41:51.000000000 +0000
@@ -1136,10 +1136,9 @@
 		serial_out(up, UART_IER, up->ier);
 
 		if (up->bugs & UART_BUG_TXEN) {
-			unsigned char lsr, iir;
+			unsigned char lsr;
 			lsr = serial_in(up, UART_LSR);
-			iir = serial_in(up, UART_IIR);
-			if (lsr & UART_LSR_TEMT && iir & UART_IIR_NO_INT)
+			if (lsr & UART_LSR_TEMT)
 				transmit_chars(up);
 		}
 	}

Re: [PATCH] IRQ handling race and spurious IIR read in serial/8250.c

[Ian Jackson]

I wrote:
> In drivers/serial/8250.c in Linux there are two bugs:
>  1. UART_BUG_TXEN can be spuriously set, due to an IRQ race
>  2. The workaround then applied by the kernel is itself buggy

Markus Armbruster has also experienced this problem in a Xen
environment and has confirmed that my patch fixes it.

I think at the very least this change:

> Proposed initial band-aid fix (against 2.6.28.4):
> 
> Do not read IIR in serial8250_start_tx when UART_BUG_TXEN
> 
> Reading the IIR clears some oustanding interrupts so it is not safe.
> Instead, simply transmit immediately if the buffer is empty without
> regard to IIR.
> 
> Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>

should be made right away.

Ian.

Re: [PATCH] IRQ handling race and spurious IIR read in serial/8250.c

[Ian Jackson]

I wrote:
> In drivers/serial/8250.c in Linux there are two bugs:
>  1. UART_BUG_TXEN can be spuriously set, due to an IRQ race
>  2. The workaround then applied by the kernel is itself buggy

Markus Armbruster has also experienced this problem in a Xen
environment and has confirmed that my patch fixes it.

I think at the very least this change:

> Proposed initial band-aid fix (against 2.6.28.4):
> 
> Do not read IIR in serial8250_start_tx when UART_BUG_TXEN
> 
> Reading the IIR clears some oustanding interrupts so it is not safe.
> Instead, simply transmit immediately if the buffer is empty without
> regard to IIR.
> 
> Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>

should be made right away.

Ian.

Re: [Xen-devel] Re: [PATCH] IRQ handling race and spurious IIR read in serial/8250.c

[Markus Armbruster]

Ian Jackson <Ian.Jackson@eu.citrix.com> writes:

> I wrote:
>> In drivers/serial/8250.c in Linux there are two bugs:
>>  1. UART_BUG_TXEN can be spuriously set, due to an IRQ race
>>  2. The workaround then applied by the kernel is itself buggy
>
> Markus Armbruster has also experienced this problem in a Xen
> environment and has confirmed that my patch fixes it.

Correct.

> I think at the very least this change:
>
>> Proposed initial band-aid fix (against 2.6.28.4):
>> 
>> Do not read IIR in serial8250_start_tx when UART_BUG_TXEN
>> 
>> Reading the IIR clears some oustanding interrupts so it is not safe.
>> Instead, simply transmit immediately if the buffer is empty without
>> regard to IIR.
>> 
>> Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>
>
> should be made right away.
>
> Ian.

Patch makes sense to me.

Acked-by: Markus Armbruster <armbru@redhat.com>

Re: Re: [PATCH] IRQ handling race and spurious IIR read in serial/8250.c

[Markus Armbruster]

Ian Jackson <Ian.Jackson@eu.citrix.com> writes:

> I wrote:
>> In drivers/serial/8250.c in Linux there are two bugs:
>>  1. UART_BUG_TXEN can be spuriously set, due to an IRQ race
>>  2. The workaround then applied by the kernel is itself buggy
>
> Markus Armbruster has also experienced this problem in a Xen
> environment and has confirmed that my patch fixes it.

Correct.

> I think at the very least this change:
>
>> Proposed initial band-aid fix (against 2.6.28.4):
>> 
>> Do not read IIR in serial8250_start_tx when UART_BUG_TXEN
>> 
>> Reading the IIR clears some oustanding interrupts so it is not safe.
>> Instead, simply transmit immediately if the buffer is empty without
>> regard to IIR.
>> 
>> Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>
>
> should be made right away.
>
> Ian.

Patch makes sense to me.

Acked-by: Markus Armbruster <armbru@redhat.com>

Re: [PATCH] IRQ handling race and spurious IIR read in serial/8250.c

[Jeremy Fitzhardinge]

Added cc:

Ian Jackson wrote:
> Anders Kaseorg writes ("Re: Serial console hangs with Linux 2.6.20 HVM guest"):
>   
>> Yes, I took Linux v2.6.20 on amd64, ran `make defconfig`, then ran `make 
>> menuconfig` and turned off CONFIG_HOTPLUG_CPU (Processor type and features 
>> ò Support for hot-pluggable CPUs).
>>     
>
> Thanks.  I think I have tracked down the bug.  In
> drivers/serial/8250.c in Linux there are two bugs:
>  1. UART_BUG_TXEN can be spuriously set, due to an IRQ race
>  2. The workaround then applied by the kernel is itself buggy
>
> Anders: can you try two tests for me ?  Firstly, in
> serial8250_startup, delete the section which sets UART_BUG_TXEN (see
> 2nd patch below); I think this will fix the symptoms for you.
> Secondly, in serial8250_start_tx delete the read from the IIR and the
> relevant branch of the text (see 3rd patch below); I think this will
> also in itself fix your symptoms.  I haven't compiled either patch (so
> you may find that eg I missed deleting some variables).
>
>
> The bugs in detail (this discussion applies to 2.6.20 and also to
> 2.6.28.4):
>
>  1. The hunk of serial8250_startup I quote below attempts to discover
>     whether writing the IER re-asserts the THRI (transmit ready)
>     interrupt.  However the spinlock that it has taken out,
>     port->lock, is not the one that the IRQ service routine takes
>     before reading the IIR (i->lock).  As a result, on an SMP system
>     the generated interrupt races with the straight-line code in
>     serial8250_startup.
>
>     If serial8250_startup loses the race (perhaps because the system
>     is a VM and its VCPU got preempted), UART_BUG_TXEN is spuriously
>     added to bugs.  This is quite unlikely in a normal system but in
>     certain Xen configurations, particularly ones where there is CPU
>     pressure, we may lose the race every time.
>
>     It is not exactly clear to me how this ought to be resolved.  One
>     possibility is that the UART_BUG_TXEN problem might be worked
>     around perfectly well by the new and very similar workaround
>     UART_BUG_THRE[1] in 2.6.21ish in which case it could just be
>     removed.
>     
>  2. UART_BUG_TXEN's workaround appears to be intended to be harmless.
>     However what it actually does is to read the IIR, thus clearing
>     any actual interrupt (including incidentally non-THRI), and then
>     only perform the intended servicing if the interrupt was _not_
>     asserted.  That is, it breaks on any serial port with the bug.
>
>     As far as I can see there is not much use in UART_BUG_TXEN reading
>     IIR at all, so a suitable change if we want to keep UART_BUG_TXEN
>     might be the first patch I enclose below (again, not compiled
>     or tested).
>
>     If UART_BUG_TXEN is retained something along these lines should be
>     done at the very least.
>
> Ian.
>
> [1] http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=40b36daad0ac704e6d5c1b75789f371ef5b053c1
>     in which case UART
>
>
> Proposed initial band-aid fix (against 2.6.28.4):
>
>
> Do not read IIR in serial8250_start_tx when UART_BUG_TXEN
>
> Reading the IIR clears some oustanding interrupts so it is not safe.
> Instead, simply transmit immediately if the buffer is empty without
> regard to IIR.
>
> Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>
>
> --- ../linux-2.6.28.4/drivers/serial/8250.c~	2009-02-06 21:47:45.000000000 +0000
> +++ ../linux-2.6.28.4/drivers/serial/8250.c	2009-02-11 15:55:24.000000000 +0000
> @@ -1257,14 +1257,12 @@
>  		serial_out(up, UART_IER, up->ier);
>  
>  		if (up->bugs & UART_BUG_TXEN) {
> -			unsigned char lsr, iir;
> +			unsigned char lsr;
>  			lsr = serial_in(up, UART_LSR);
>  			up->lsr_saved_flags |= lsr & LSR_SAVE_FLAGS;
> -			iir = serial_in(up, UART_IIR) & 0x0f;
>  			if ((up->port.type == PORT_RM9000) ?
> -				(lsr & UART_LSR_THRE &&
> -				(iir == UART_IIR_NO_INT || iir == UART_IIR_THRI)) :
> -				(lsr & UART_LSR_TEMT && iir & UART_IIR_NO_INT))
> +				(lsr & UART_LSR_THRE) :
> +				(lsr & UART_LSR_TEMT))
>  				transmit_chars(up);
>  		}
>  	}
>
>
> Anders - first patch to try (against 2.6.20):
>
> --- drivers/serial/8250.c~	2007-02-04 18:44:54.000000000 +0000
> +++ drivers/serial/8250.c	2009-02-11 15:39:43.000000000 +0000
> @@ -1645,25 +1645,6 @@
>  
>  	serial8250_set_mctrl(&up->port, up->port.mctrl);
>  
> -	/*
> -	 * Do a quick test to see if we receive an
> -	 * interrupt when we enable the TX irq.
> -	 */
> -	serial_outp(up, UART_IER, UART_IER_THRI);
> -	lsr = serial_in(up, UART_LSR);
> -	iir = serial_in(up, UART_IIR);
> -	serial_outp(up, UART_IER, 0);
> -
> -	if (lsr & UART_LSR_TEMT && iir & UART_IIR_NO_INT) {
> -		if (!(up->bugs & UART_BUG_TXEN)) {
> -			up->bugs |= UART_BUG_TXEN;
> -			pr_debug("ttyS%d - enabling bad tx status workarounds\n",
> -				 port->line);
> -		}
> -	} else {
> -		up->bugs &= ~UART_BUG_TXEN;
> -	}
> -
>  	spin_unlock_irqrestore(&up->port.lock, flags);
>  
>  	/*
>
>
> Anders - second patch to try (against 2.6.20):
> Fix should be suitable for distribution IMO.
>
> Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>
>
> --- drivers/serial/8250.c~	2007-02-04 18:44:54.000000000 +0000
> +++ drivers/serial/8250.c	2009-02-11 15:41:51.000000000 +0000
> @@ -1136,10 +1136,9 @@
>  		serial_out(up, UART_IER, up->ier);
>  
>  		if (up->bugs & UART_BUG_TXEN) {
> -			unsigned char lsr, iir;
> +			unsigned char lsr;
>  			lsr = serial_in(up, UART_LSR);
> -			iir = serial_in(up, UART_IIR);
> -			if (lsr & UART_LSR_TEMT && iir & UART_IIR_NO_INT)
> +			if (lsr & UART_LSR_TEMT)
>  				transmit_chars(up);
>  		}
>  	}
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/
>   

Re: [PATCH] IRQ handling race and spurious IIR read in serial/8250.c

[Ian Jackson]

Alan Cox writes ("Re: [PATCH] IRQ handling race and spurious IIR read in serial/8250.c"):
> > Ian Jackson recently debugged a problem reported by Anders Kaseorg, and
> > posted a fix (see http://lkml.org/lkml/2009/2/11/240).  As far as I can
> > see, the patch has been dropped on the floor.
> 
> It was proposed as a band aid fix not a real fix. The real fix involves
> finishing sorting out any locking issues Ian identified. I've not seen a
> patch for that yet.

The patch I have provided and which Markus has reposted has these
properties:

 1. It is definitely correct and does not introduce any new bugs.

 2. It makes the existing bug definitely go away.  That is, after my
    patch is applied the code drives the UART correctly - that is,
    according to the specification and in a manner guaranteed to be
    reliable - even though the code may still mistakenly set a bug
    flag;

 3. Without it the code is definitely wrong;

 4. Any fix which does not involve completely removing UART_BUG_TXEN
    will need my change.   

Or to put it another way

                        UART_BUG_TXEN           UART_BUG_TXEN
                        incorrectly set         behaviour

    Current code         Sometimes in VMs        Breaks correct systems
                         Rarely in baremetal?    (observed in production)

    With my patch        Sometimes in VMs        Always correct.
                         (behaviour unchanged)   Minor possible
                                                  performance problem.

    Perfect code         Perhaps feature         If retained, should
                         should be removed?       be correct.  Minor
                                                  performance impact
                                                  is acceptable.

My patch is therefore a strict improvement and also a likely component
of many of the possible ultimate fixes; the other ultimate fixes
involve removing entirely the code which I am now proposing to patch.

Please do not block this bugfix just because I haven't been able to
conclusively determine whether the buggy-without-my-patch workaround
should be entirely removed, and just because I do not fix every other
bug in the same area.

In particular, the fact that the detection of UART_BUG_TXEN remains
buggy is not a reason to block a patch which makes UART_BUG_TXEN's
effects correct.

As you can see `perfect code' is not yet attainable because the people
who originally implemented UART_BUG_TXEN don't seem to be around right
now to ask, and we don't have a 100% clear description of the buggy
behaviour it is trying to detect, and so we don't know whether it can
safely be removed.

Ian.

Re: [PATCH] IRQ handling race and spurious IIR read in serial/8250.c

[Ian Jackson]

Alan Cox writes ("Re: [PATCH] IRQ handling race and spurious IIR read in serial/8250.c"):
> > Ian Jackson recently debugged a problem reported by Anders Kaseorg, and
> > posted a fix (see http://lkml.org/lkml/2009/2/11/240).  As far as I can
> > see, the patch has been dropped on the floor.
> 
> It was proposed as a band aid fix not a real fix. The real fix involves
> finishing sorting out any locking issues Ian identified. I've not seen a
> patch for that yet.

The patch I have provided and which Markus has reposted has these
properties:

 1. It is definitely correct and does not introduce any new bugs.

 2. It makes the existing bug definitely go away.  That is, after my
    patch is applied the code drives the UART correctly - that is,
    according to the specification and in a manner guaranteed to be
    reliable - even though the code may still mistakenly set a bug
    flag;

 3. Without it the code is definitely wrong;

 4. Any fix which does not involve completely removing UART_BUG_TXEN
    will need my change.   

Or to put it another way

                        UART_BUG_TXEN           UART_BUG_TXEN
                        incorrectly set         behaviour

    Current code         Sometimes in VMs        Breaks correct systems
                         Rarely in baremetal?    (observed in production)

    With my patch        Sometimes in VMs        Always correct.
                         (behaviour unchanged)   Minor possible
                                                  performance problem.

    Perfect code         Perhaps feature         If retained, should
                         should be removed?       be correct.  Minor
                                                  performance impact
                                                  is acceptable.

My patch is therefore a strict improvement and also a likely component
of many of the possible ultimate fixes; the other ultimate fixes
involve removing entirely the code which I am now proposing to patch.

Please do not block this bugfix just because I haven't been able to
conclusively determine whether the buggy-without-my-patch workaround
should be entirely removed, and just because I do not fix every other
bug in the same area.

In particular, the fact that the detection of UART_BUG_TXEN remains
buggy is not a reason to block a patch which makes UART_BUG_TXEN's
effects correct.

As you can see `perfect code' is not yet attainable because the people
who originally implemented UART_BUG_TXEN don't seem to be around right
now to ask, and we don't have a 100% clear description of the buggy
behaviour it is trying to detect, and so we don't know whether it can
safely be removed.

Ian.

Re: [PATCH] IRQ handling race and spurious IIR read in serial/8250.c

[Alan Cox]

> Ian Jackson recently debugged a problem reported by Anders Kaseorg, and
> posted a fix (see http://lkml.org/lkml/2009/2/11/240).  As far as I can
> see, the patch has been dropped on the floor.

It was proposed as a band aid fix not a real fix. The real fix involves
finishing sorting out any locking issues Ian identified. I've not seen a
patch for that yet.

Re: [PATCH] IRQ handling race and spurious IIR read in serial/8250.c

[Alan Cox]

> Ian Jackson recently debugged a problem reported by Anders Kaseorg, and
> posted a fix (see http://lkml.org/lkml/2009/2/11/240).  As far as I can
> see, the patch has been dropped on the floor.

It was proposed as a band aid fix not a real fix. The real fix involves
finishing sorting out any locking issues Ian identified. I've not seen a
patch for that yet.

[PATCH] IRQ handling race and spurious IIR read in serial/8250.c

[Markus Armbruster]

From: Ian Jackson <ian.jackson@eu.citrix.com>

Do not read IIR in serial8250_start_tx when UART_BUG_TXEN

Reading the IIR clears some oustanding interrupts so it is not safe.
Instead, simply transmit immediately if the buffer is empty without
regard to IIR.

Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>
Reviewed-by: Markus Armbruster <armbru@redhat.com>

---
Ian Jackson recently debugged a problem reported by Anders Kaseorg, and
posted a fix (see http://lkml.org/lkml/2009/2/11/240).  As far as I can
see, the patch has been dropped on the floor.

I also experienced the problem, and Ian's patch fixes it for me.

The bug bites Xen HVM guests.  Output to the serial console stalls until
some input happens.  As Ian's analysis quoted below shows, it is a race
condition that could theoretically bite elsewhere as well.

Ian Jackson explained:
> The bugs in detail (this discussion applies to 2.6.20 and also to
> 2.6.28.4):
>
>  1. The hunk of serial8250_startup I quote below attempts to discover
>     whether writing the IER re-asserts the THRI (transmit ready)
>     interrupt.  However the spinlock that it has taken out,
>     port->lock, is not the one that the IRQ service routine takes
>     before reading the IIR (i->lock).  As a result, on an SMP system
>     the generated interrupt races with the straight-line code in
>     serial8250_startup.
>
>     If serial8250_startup loses the race (perhaps because the system
>     is a VM and its VCPU got preempted), UART_BUG_TXEN is spuriously
>     added to bugs.  This is quite unlikely in a normal system but in
>     certain Xen configurations, particularly ones where there is CPU
>     pressure, we may lose the race every time.
>
>     It is not exactly clear to me how this ought to be resolved.  One
>     possibility is that the UART_BUG_TXEN problem might be worked
>     around perfectly well by the new and very similar workaround
>     UART_BUG_THRE[1] in 2.6.21ish in which case it could just be
>     removed.
>      2. UART_BUG_TXEN's workaround appears to be intended to be
> harmless.
>     However what it actually does is to read the IIR, thus clearing
>     any actual interrupt (including incidentally non-THRI), and then
>     only perform the intended servicing if the interrupt was _not_
>     asserted.  That is, it breaks on any serial port with the bug.
>
>     As far as I can see there is not much use in UART_BUG_TXEN reading
>     IIR at all, so a suitable change if we want to keep UART_BUG_TXEN
>     might be the first patch I enclose below (again, not compiled
>     or tested).
>
>     If UART_BUG_TXEN is retained something along these lines should be
>     done at the very least.
>
> Ian.
>
> [1] http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=40b36daad0ac704e6d5c1b75789f371ef5b053c1
>     in which case UART

--- ../linux-2.6.28.4/drivers/serial/8250.c~	2009-02-06 21:47:45.000000000 +0000
+++ ../linux-2.6.28.4/drivers/serial/8250.c	2009-02-11 15:55:24.000000000 +0000
@@ -1257,14 +1257,12 @@
 		serial_out(up, UART_IER, up->ier);
 
 		if (up->bugs & UART_BUG_TXEN) {
-			unsigned char lsr, iir;
+			unsigned char lsr;
 			lsr = serial_in(up, UART_LSR);
 			up->lsr_saved_flags |= lsr & LSR_SAVE_FLAGS;
-			iir = serial_in(up, UART_IIR) & 0x0f;
 			if ((up->port.type == PORT_RM9000) ?
-				(lsr & UART_LSR_THRE &&
-				(iir == UART_IIR_NO_INT || iir == UART_IIR_THRI)) :
-				(lsr & UART_LSR_TEMT && iir & UART_IIR_NO_INT))
+				(lsr & UART_LSR_THRE) :
+				(lsr & UART_LSR_TEMT))
 				transmit_chars(up);
 		}
 	}

[PATCH] IRQ handling race and spurious IIR read in serial/8250.c

[Markus Armbruster]

From: Ian Jackson <ian.jackson@eu.citrix.com>

Do not read IIR in serial8250_start_tx when UART_BUG_TXEN

Reading the IIR clears some oustanding interrupts so it is not safe.
Instead, simply transmit immediately if the buffer is empty without
regard to IIR.

Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>
Reviewed-by: Markus Armbruster <armbru@redhat.com>

---
Ian Jackson recently debugged a problem reported by Anders Kaseorg, and
posted a fix (see http://lkml.org/lkml/2009/2/11/240).  As far as I can
see, the patch has been dropped on the floor.

I also experienced the problem, and Ian's patch fixes it for me.

The bug bites Xen HVM guests.  Output to the serial console stalls until
some input happens.  As Ian's analysis quoted below shows, it is a race
condition that could theoretically bite elsewhere as well.

Ian Jackson explained:
> The bugs in detail (this discussion applies to 2.6.20 and also to
> 2.6.28.4):
>
>  1. The hunk of serial8250_startup I quote below attempts to discover
>     whether writing the IER re-asserts the THRI (transmit ready)
>     interrupt.  However the spinlock that it has taken out,
>     port->lock, is not the one that the IRQ service routine takes
>     before reading the IIR (i->lock).  As a result, on an SMP system
>     the generated interrupt races with the straight-line code in
>     serial8250_startup.
>
>     If serial8250_startup loses the race (perhaps because the system
>     is a VM and its VCPU got preempted), UART_BUG_TXEN is spuriously
>     added to bugs.  This is quite unlikely in a normal system but in
>     certain Xen configurations, particularly ones where there is CPU
>     pressure, we may lose the race every time.
>
>     It is not exactly clear to me how this ought to be resolved.  One
>     possibility is that the UART_BUG_TXEN problem might be worked
>     around perfectly well by the new and very similar workaround
>     UART_BUG_THRE[1] in 2.6.21ish in which case it could just be
>     removed.
>      2. UART_BUG_TXEN's workaround appears to be intended to be
> harmless.
>     However what it actually does is to read the IIR, thus clearing
>     any actual interrupt (including incidentally non-THRI), and then
>     only perform the intended servicing if the interrupt was _not_
>     asserted.  That is, it breaks on any serial port with the bug.
>
>     As far as I can see there is not much use in UART_BUG_TXEN reading
>     IIR at all, so a suitable change if we want to keep UART_BUG_TXEN
>     might be the first patch I enclose below (again, not compiled
>     or tested).
>
>     If UART_BUG_TXEN is retained something along these lines should be
>     done at the very least.
>
> Ian.
>
> [1] http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=40b36daad0ac704e6d5c1b75789f371ef5b053c1
>     in which case UART

--- ../linux-2.6.28.4/drivers/serial/8250.c~	2009-02-06 21:47:45.000000000 +0000
+++ ../linux-2.6.28.4/drivers/serial/8250.c	2009-02-11 15:55:24.000000000 +0000
@@ -1257,14 +1257,12 @@
 		serial_out(up, UART_IER, up->ier);
 
 		if (up->bugs & UART_BUG_TXEN) {
-			unsigned char lsr, iir;
+			unsigned char lsr;
 			lsr = serial_in(up, UART_LSR);
 			up->lsr_saved_flags |= lsr & LSR_SAVE_FLAGS;
-			iir = serial_in(up, UART_IIR) & 0x0f;
 			if ((up->port.type == PORT_RM9000) ?
-				(lsr & UART_LSR_THRE &&
-				(iir == UART_IIR_NO_INT || iir == UART_IIR_THRI)) :
-				(lsr & UART_LSR_TEMT && iir & UART_IIR_NO_INT))
+				(lsr & UART_LSR_THRE) :
+				(lsr & UART_LSR_TEMT))
 				transmit_chars(up);
 		}
 	}