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

[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: [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: [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/
>   

[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);
 		}
 	}

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

[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.