[Qemu-devel] Re: 2.6.24 says "serial8250: too much work for irq4" a lot.

[Qemu-devel] Re: 2.6.24 says "serial8250: too much work for irq4" a lot.

[Rob Landley]

On Tuesday 05 February 2008 15:07:27 H. Peter Anvin wrote:
> Rob Landley wrote:
> > When running a 2.6.24 kernel built for x86-64 under qemu via serial
> > console, doing CPU-intensive things that also produce a lot of output
> > (such as compiling software) tends to produce the error message in the
> > title.
> >
> > Anybody have a clue why?  It doesn't seem to cause an actual problem, but
> > it's kind of annoying.
> >
> > (If it's a qemu issue, I can go bother them.  It's possible that qemu
> > isn't delivering interrupts as often as it expects, since that's limited
> > by the granularity of the host timer; I know the clock in qemu can run a
> > bit slow because it only gets clock interrupts when the host system isn't
> > too busy to schedule the emulator.  But this doesn't usually cause a
> > problem.  I _think_ the message is just a "this should never happen" type
> > warning, which is happening to me.  But I break stuff. :)
>
> This is because Qemu spews data to the serial port without any rate
> limiting; this causes the in-kernel serial port driver to think the port
> is stuck.  The serial port emulation needs to make it possible to drain
> the virtual FIFO every now and then, as opposed to filling it again
> immediately.
>
> 	-hpa

Thanks.  cc'd the right list on this one. :)

Rob
-- 
"One of my most productive days was throwing away 1000 lines of code."
  - Ken Thompson.

[Qemu-devel] Re: 2.6.24 says "serial8250: too much work for irq4" a lot.

[Rob Landley]

Here's a patch Peter Anvin wrote so the serial I/O doesn't flood the kernel.

Here's the thread on linux-kernel aboout it:
http://lkml.org/lkml/2008/2/5/401

Rob

On Thursday 07 February 2008 15:19:39 you wrote:
> Rob Landley wrote:
> > Specifically, qemu isn't paravirtualized, it's fully virtualized.  The
> > same kernel can run on real hardware just fine.  (Sort of the point of
> > the project...)
> >
> > I can yank the warning for the kernels I build (or set PASS_LIMIT to
> > 9999999), but I'd rather not carry any more patches than I can avoid...
>
> Patch attached, completely untested beyond compilation.
>
> In particular:
>
> 	- we should probably clear the burst counter when the interrupt
> 	  line goes from inactive to active.
> 	- there probably should be a timer which clears the burst
>   	  counter.
>
> However, I think I've covered most of the bases...
>
> 	-hpa

diff --git a/hw/serial.c b/hw/serial.c
index b1bd0ff..c902792 100644
--- a/hw/serial.c
+++ b/hw/serial.c
@@ -73,6 +73,15 @@
 #define UART_LSR_OE	0x02	/* Overrun error indicator */
 #define UART_LSR_DR	0x01	/* Receiver data ready */
 
+/*
+ * It's common for an IRQ handler to keep reading the RBR until
+ * the LSR indicates that the FIFO is empty, expecting that the
+ * CPU is vastly faster than the serial line.  This can cause
+ * overruns or error indications if the FIFO never empties, so
+ * give the target OS a breather every so often.
+ */
+#define MAX_BURST	512
+
 struct SerialState {
     uint16_t divider;
     uint8_t rbr; /* receive register */
@@ -91,8 +100,14 @@ struct SerialState {
     int last_break_enable;
     target_phys_addr_t base;
     int it_shift;
+    int burst_len;
 };
 
+static void serial_clear_burst(SerialState *s)
+{
+    s->burst_len = 0;
+}
+
 static void serial_update_irq(SerialState *s)
 {
     if ((s->lsr & UART_LSR_DR) && (s->ier & UART_IER_RDI)) {
@@ -114,6 +129,8 @@ static void serial_update_parameters(SerialState *s)
     int speed, parity, data_bits, stop_bits;
     QEMUSerialSetParams ssp;
 
+    serial_clear_burst(s);
+
     if (s->lcr & 0x08) {
         if (s->lcr & 0x10)
             parity = 'E';
@@ -221,9 +238,12 @@ static uint32_t serial_ioport_read(void *opaque, uint32_t addr)
             ret = s->divider & 0xff;
         } else {
             ret = s->rbr;
-            s->lsr &= ~(UART_LSR_DR | UART_LSR_BI);
-            serial_update_irq(s);
-            qemu_chr_accept_input(s->chr);
+	    if (s->burst_len < MAX_BURST) {
+		s->burst_len++;
+		s->lsr &= ~(UART_LSR_DR | UART_LSR_BI);
+		serial_update_irq(s);
+		qemu_chr_accept_input(s->chr);
+	    }
         }
         break;
     case 1:
@@ -235,6 +255,7 @@ static uint32_t serial_ioport_read(void *opaque, uint32_t addr)
         break;
     case 2:
         ret = s->iir;
+	serial_clear_burst(s);
         /* reset THR pending bit */
         if ((ret & 0x7) == UART_IIR_THRI)
             s->thr_ipending = 0;
@@ -248,6 +269,10 @@ static uint32_t serial_ioport_read(void *opaque, uint32_t addr)
         break;
     case 5:
         ret = s->lsr;
+	if (s->burst_len >= MAX_BURST)
+	    ret &= ~(UART_LSR_DR|UART_LSR_BI);
+	if (!(ret & UART_LSR_DR))
+	    serial_clear_burst(s);
         break;
     case 6:
         if (s->mcr & UART_MCR_LOOP) {


-- 
"One of my most productive days was throwing away 1000 lines of code."
  - Ken Thompson.

Re: [Qemu-devel] Re: 2.6.24 says "serial8250: too much work for irq4" a lot.

[Blue Swirl]

On 2/9/08, Rob Landley <rob@landley.net> wrote:
> Here's a patch Peter Anvin wrote so the serial I/O doesn't flood the kernel.

The patch looks OK, but the throttling should benefit all devices, as
discussed here:
http://lists.gnu.org/archive/html/qemu-devel/2007-12/msg00283.html

Re: [Qemu-devel] Re: 2.6.24 says "serial8250: too much work for irq4" a lot.

[H. Peter Anvin]

Blue Swirl wrote:
> On 2/9/08, Rob Landley <rob@landley.net> wrote:
>> Here's a patch Peter Anvin wrote so the serial I/O doesn't flood the kernel.
> 
> The patch looks OK, but the throttling should benefit all devices, as
> discussed here:
> http://lists.gnu.org/archive/html/qemu-devel/2007-12/msg00283.html

I strongly disagree with the sentiments in that post.

This is not a matter of rate throttling, but simulated FIFO exhaustion 
-- they are NOT the same thing.  Simulated FIFO exhaustion is 
functionally equivalent to making sure there are interrupt windows 
opened in an otherwise-too-long critical section; it doesn't constrain 
any particular flow rate, as it still permits another interrupt to 
immediately come in.

If you look at the patch, there are no timing dependencies; the only 
parameter is the depth of the virtual queue.  The exhaustion is 
completely controlled by target OS access patterns.

	-hpa

Re: [Qemu-devel] Re: 2.6.24 says "serial8250: too much work for irq4" a lot.

[Blue Swirl]

On 2/9/08, H. Peter Anvin <hpa@zytor.com> wrote:
> Blue Swirl wrote:
> > On 2/9/08, Rob Landley <rob@landley.net> wrote:
> >> Here's a patch Peter Anvin wrote so the serial I/O doesn't flood the kernel.
> >
> > The patch looks OK, but the throttling should benefit all devices, as
> > discussed here:
> > http://lists.gnu.org/archive/html/qemu-devel/2007-12/msg00283.html
>
> I strongly disagree with the sentiments in that post.
>
> This is not a matter of rate throttling, but simulated FIFO exhaustion
> -- they are NOT the same thing.  Simulated FIFO exhaustion is
> functionally equivalent to making sure there are interrupt windows
> opened in an otherwise-too-long critical section; it doesn't constrain
> any particular flow rate, as it still permits another interrupt to
> immediately come in.
>
> If you look at the patch, there are no timing dependencies; the only
> parameter is the depth of the virtual queue.  The exhaustion is
> completely controlled by target OS access patterns.

Thanks, this clarified the difference. But I'll rephrase my original comment:

The patch looks OK, but the simulated FIFO exhaustion should benefit
all devices, as
discussed here:
http://lists.gnu.org/archive/html/qemu-devel/2007-12/msg00283.html

Re: [Qemu-devel] Re: 2.6.24 says "serial8250: too much work for irq4" a lot.

[H. Peter Anvin]

Blue Swirl wrote:
>>
>> If you look at the patch, there are no timing dependencies; the only
>> parameter is the depth of the virtual queue.  The exhaustion is
>> completely controlled by target OS access patterns.
> 
> Thanks, this clarified the difference. But I'll rephrase my original comment:
> 
> The patch looks OK, but the simulated FIFO exhaustion should benefit
> all devices, as
> discussed here:
> http://lists.gnu.org/archive/html/qemu-devel/2007-12/msg00283.html

The difference is you *can't* do that in a general layer.

	-hpa

Re: [Qemu-devel] Re: 2.6.24 says "serial8250: too much work for irq4" a lot.

[Blue Swirl]

On 2/9/08, H. Peter Anvin <hpa@zytor.com> wrote:
> Blue Swirl wrote:
> >>
> >> If you look at the patch, there are no timing dependencies; the only
> >> parameter is the depth of the virtual queue.  The exhaustion is
> >> completely controlled by target OS access patterns.
> >
> > Thanks, this clarified the difference. But I'll rephrase my original comment:
> >
> > The patch looks OK, but the simulated FIFO exhaustion should benefit
> > all devices, as
> > discussed here:
> > http://lists.gnu.org/archive/html/qemu-devel/2007-12/msg00283.html
>
> The difference is you *can't* do that in a general layer.

What makes you think that is impossible? Just move the
serial_clear_burst to vl.c under name chr_clear_burst, move burst_len
to CharDriverState and introduce a new function in vl.c that contains
the burst length check. This is functionally identical to your patch.
For 100% compatibility, the init functions could be changed so that
only PC serial is affected, but I think all character devices would
benefit from this.

Also win2k install hack in ide.c seems to be related to this problem,
so even more generic solution would be desirable.

Re: [Qemu-devel] Re: 2.6.24 says "serial8250: too much work for irq4" a lot.

[Paul Brook]

On Sunday 10 February 2008, Blue Swirl wrote:
> On 2/9/08, H. Peter Anvin <hpa@zytor.com> wrote:
> > Blue Swirl wrote:
> > >> If you look at the patch, there are no timing dependencies; the only
> > >> parameter is the depth of the virtual queue.  The exhaustion is
> > >> completely controlled by target OS access patterns.
> > >
> > > Thanks, this clarified the difference. But I'll rephrase my original
> > > comment:
> > >
> > > The patch looks OK, but the simulated FIFO exhaustion should benefit
> > > all devices, as
> > > discussed here:
> > > http://lists.gnu.org/archive/html/qemu-devel/2007-12/msg00283.html
> >
> > The difference is you *can't* do that in a general layer.
>
> What makes you think that is impossible? 

IIUC the proposed patch makes the serial driver return an empty FIFO exactly 
once, them immediately continue receiving data. Throughput should be 
approximately the same, you've just got a bit of extra overhead to process 
the additional interrupts.  This is very different to the previous patch 
which did time-based throughput limiting.

You can't do this in generic code because there's no way to guess when the 
guest os has seen the FIFO empty condition. The best you can do is pause for 
some arbitrary length of time, which is both unreliable (the guest OS may not 
have got to far enough yet, especially if the host machine is heavily 
loaded), and has a significant negative impact on throughput.

> Also win2k install hack in ide.c seems to be related to this problem,
> so even more generic solution would be desirable.

IIUC the win2k hack is an actual timing problem. The win2k IDE drivers are 
buggy, and fall over if the drive responds too soon.

Paul

Re: [Qemu-devel] Re: 2.6.24 says "serial8250: too much work for irq4" a lot.

[Aurelien Jarno]

On Sun, Feb 10, 2008 at 11:26:06AM +0000, Paul Brook wrote:
> On Sunday 10 February 2008, Blue Swirl wrote:
> > On 2/9/08, H. Peter Anvin <hpa@zytor.com> wrote:
> > > Blue Swirl wrote:
> > > >> If you look at the patch, there are no timing dependencies; the only
> > > >> parameter is the depth of the virtual queue.  The exhaustion is
> > > >> completely controlled by target OS access patterns.
> > > >
> > > > Thanks, this clarified the difference. But I'll rephrase my original
> > > > comment:
> > > >
> > > > The patch looks OK, but the simulated FIFO exhaustion should benefit
> > > > all devices, as
> > > > discussed here:
> > > > http://lists.gnu.org/archive/html/qemu-devel/2007-12/msg00283.html
> > >
> > > The difference is you *can't* do that in a general layer.
> >
> > What makes you think that is impossible? 
> 
> IIUC the proposed patch makes the serial driver return an empty FIFO exactly 
> once, them immediately continue receiving data. Throughput should be 
> approximately the same, you've just got a bit of extra overhead to process 
> the additional interrupts.  This is very different to the previous patch 
> which did time-based throughput limiting.
> 
> You can't do this in generic code because there's no way to guess when the 
> guest os has seen the FIFO empty condition. The best you can do is pause for 
> some arbitrary length of time, which is both unreliable (the guest OS may not 
> have got to far enough yet, especially if the host machine is heavily 
> loaded), and has a significant negative impact on throughput.
> 
> > Also win2k install hack in ide.c seems to be related to this problem,
> > so even more generic solution would be desirable.
> 
> IIUC the win2k hack is an actual timing problem. The win2k IDE drivers are 
> buggy, and fall over if the drive responds too soon.
> 

Is everybody convinced about this patch now? I would really like to see
it in the CVS, as it greatly improves the usability of the serial 
consoles on targets running GNU/Linux (and probably OSes).

-- 
  .''`.  Aurelien Jarno	            | GPG: 1024D/F1BCDB73
 : :' :  Debian developer           | Electrical Engineer
 `. `'   aurel32@debian.org         | aurelien@aurel32.net
   `-    people.debian.org/~aurel32 | www.aurel32.net