Re: [PATCH] 2.5.68 Fix IO_APIC IRQ assignment bug

Re: [PATCH] 2.5.68 Fix IO_APIC IRQ assignment bug

[Chuck Ebbert]

> Yes, we need to bail out in assign_irq_vector when we wrap around, 
> otherwise we cause collisions when programming the IOAPIC. And we also 
> need to avoid overruning NR_IRQS structures in setup_IO_APIC_irqs.


  Do you mean the panic on running out of sources should be put
back in?



------
 Chuck

Re: [PATCH] 2.5.68 Fix IO_APIC IRQ assignment bug

[Zwane Mwaikambo]

On Mon, 21 Apr 2003, Chuck Ebbert wrote:

> > Yes, we need to bail out in assign_irq_vector when we wrap around, 
> > otherwise we cause collisions when programming the IOAPIC. And we also 
> > need to avoid overruning NR_IRQS structures in setup_IO_APIC_irqs.
> 
> 
>   Do you mean the panic on running out of sources should be put
> back in?

No, something like this, although i think Linus hates something about it. 
Note i forgot to subtract 0x80 from the total vector count in the email 
(should be NR_IRQ_VECTORS=189).

Yes i have a system which oopses due to the high irq count with mainline. 
The NR_IRQS overrun is one, the vector collisions is another.

Date: Mon, 7 Apr 2003 22:04:13 -0400 (EDT)
From: Zwane Mwaikambo <zwane@linuxpower.ca>
X-X-Sender: zwane@montezuma.mastecende.com
To: Linus Torvalds <torvalds@transmeta.com>
cc: Linux Kernel <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH][2.5] avoid scribbling in IDT with high interrupt count.
In-Reply-To: <Pine.LNX.4.44.0304070818340.26364-100000@home.transmeta.com>
Message-ID: <Pine.LNX.4.50.0304071958360.21025-100000@montezuma.mastecende.com>
References: <Pine.LNX.4.44.0304070818340.26364-100000@home.transmeta.com>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII

On Mon, 7 Apr 2003, Linus Torvalds wrote:

> Zwane - is there any reason we couldn't just start re-using irq vector
> offsets when this happens? We already re-use the vectors themselves, so 
> restarting the offset pointer shouldn't really _change_ anything.

My patch skips irq numbers > NR_IRQS and simply return error when we run 
out of vectors, although mainline doesn't assign duplicates and cause 
collisions, it does waste vector space. e.g. for a system with 
NR_IRQS = 224 we have 23 vectors free, 0 collisions and 167 useable irqs 
and assign_irq_vectors states that we're out of vectors.

> In other words, I'm wondering if this simpler patch wouldn't be sufficient 
> instead?
> 
> Can you please test this, and re-submit (and if you can explain why your 
> patch is better, please do so - I have nothing fundamentally against it, I 
> just want to understand _why_ the complexity is needed).

Your patch booted the system but there are vector collisions resulting in 
lost irq routing when we program the IOAPIC with the duplicated vector. So 
with your patch and NR_IRQS = 224 we have 1 vectors free (0x80), 34 collisions 
and 225 irqs. However that isn't a fault of your patch but a fault with 
the NR_IRQS definition. This is what the vector space looks like on i386 
at present.

0________0x31__________________________0xef____________0xff
  system	   io interrupts            resvd vectors

0xef - 0x31 = 190 useable io interrupt vectors

So perhaps we should, apply your patch, add a NR_IRQ_VECTORS define 
and also add commentary in irq_vectors.h how does the following look? I 
had to readd the NR_IRQS checks to protect against overrunning NR_IRQS sized 
arrays and i added nr_assigned to track how many vectors were allocated so 
taht we can bail out when we're out.

Patch has been tested on a 320 interrupt system and had a maximum useable 
irq line of 211 (ethernet).

Thanks,
	Zwane

Index: linux-2.5.67/include/asm-i386/mach-default/irq_vectors.h
===================================================================
RCS file: /build/cvsroot/linux-2.5.67/include/asm-i386/mach-default/irq_vectors.h,v
retrieving revision 1.1.1.1
diff -u -p -B -r1.1.1.1 irq_vectors.h
--- linux-2.5.67/include/asm-i386/mach-default/irq_vectors.h	8 Apr 2003 01:15:29 -0000	1.1.1.1
+++ linux-2.5.67/include/asm-i386/mach-default/irq_vectors.h	8 Apr 2003 01:34:54 -0000
@@ -68,15 +68,22 @@
 #define TIMER_IRQ 0
 
 /*
- * 16 8259A IRQ's, 208 potential APIC interrupt sources.
+ * 16 8259A IRQ's, MAX_IRQ_SOURCES-16 potential APIC interrupt sources.
  * Right now the APIC is mostly only used for SMP.
  * 256 vectors is an architectural limit. (we can have
  * more than 256 devices theoretically, but they will
  * have to use shared interrupts)
  * Since vectors 0x00-0x1f are used/reserved for the CPU,
- * the usable vector space is 0x20-0xff (224 vectors)
+ * the usable vector space is 0x20-0xff (224 vectors).
+ * Linux currently makes 189 vectors available for io interrupts
+ * starting at FIRST_DEVICE_VECTOR till FIRST_SYSTEM_VECTOR
+ * 
+ * 0________0x31__________________________0xef______0xff
+ *   system           io interrupts           resvd
+ *
  */
 #ifdef CONFIG_X86_IO_APIC
+#define NR_IRQ_VECTORS	189
 #define NR_IRQS 224
 #else
 #define NR_IRQS 16
Index: linux-2.5.67/arch/i386/kernel/io_apic.c
===================================================================
RCS file: /build/cvsroot/linux-2.5.67/arch/i386/kernel/io_apic.c,v
retrieving revision 1.1.1.1
diff -u -p -B -r1.1.1.1 io_apic.c
--- linux-2.5.67/arch/i386/kernel/io_apic.c	8 Apr 2003 01:15:35 -0000	1.1.1.1
+++ linux-2.5.67/arch/i386/kernel/io_apic.c	8 Apr 2003 01:36:06 -0000
@@ -1107,9 +1107,15 @@ int irq_vector[NR_IRQS] = { FIRST_DEVICE
 
 static int __init assign_irq_vector(int irq)
 {
-	static int current_vector = FIRST_DEVICE_VECTOR, offset = 0;
+	static int current_vector = FIRST_DEVICE_VECTOR, offset = 0,
+		nr_assigned = 1;
+
 	if (IO_APIC_VECTOR(irq) > 0)
 		return IO_APIC_VECTOR(irq);
+
+	if (++nr_assigned > NR_IRQ_VECTORS)
+		return -ENOSPC;
+
 next:
 	current_vector += 8;
 	if (current_vector == SYSCALL_VECTOR)
@@ -1167,6 +1173,8 @@ void __init setup_IO_APIC_irqs(void)
 		}
 
 		irq = pin_2_irq(idx, apic, pin);
+		if (irq >= NR_IRQS)
+			continue;
 		/*
 		 * skip adding the timer int on secondary nodes, which causes
 		 * a small but painful rift in the time-space continuum
@@ -1181,6 +1189,9 @@ void __init setup_IO_APIC_irqs(void)
 
 		if (IO_APIC_IRQ(irq)) {
 			vector = assign_irq_vector(irq);
+			if (vector < 0)
+				continue;
+
 			entry.vector = vector;
 
 			if (IO_APIC_irq_trigger(irq))
@@ -2277,6 +2288,10 @@ int io_apic_set_pci_routing (int ioapic,
 {
 	struct IO_APIC_route_entry entry;
 	unsigned long flags;
+	int vector;
+
+	if (irq >= NR_IRQS)
+		return -ENOSPC;
 
 	if (!IO_APIC_IRQ(irq)) {
 		printk(KERN_ERR "IOAPIC[%d]: Invalid reference to IRQ 0/n", 
@@ -2301,8 +2316,11 @@ int io_apic_set_pci_routing (int ioapic,
 
 	add_pin_to_irq(irq, ioapic, pin);
 
-	entry.vector = assign_irq_vector(irq);
+	vector = assign_irq_vector(irq);
+	if (vector < 0)
+		return -ENOSPC;
 
+	entry.vector = vector;
 	printk(KERN_DEBUG "IOAPIC[%d]: Set PCI routing entry (%d-%d -> 0x%x -> "
 		"IRQ %d)\n", ioapic, 
 		mp_ioapics[ioapic].mpc_apicid, pin, entry.vector, irq);

-- 
function.linuxpower.ca

Re: [PATCH] 2.5.68 Fix IO_APIC IRQ assignment bug

[Mika Penttilä]

Why can't we use the same vector for multiple ioapic entrys? After all, 
we are already sharing irqs, and an irq is just a cookie for a vector. 
What do you mean with "lost irq routing" ?

--Mika


Zwane Mwaikambo wrote:

>On Mon, 21 Apr 2003, Chuck Ebbert wrote:
>
>  
>
>>>Yes, we need to bail out in assign_irq_vector when we wrap around, 
>>>otherwise we cause collisions when programming the IOAPIC. And we also 
>>>need to avoid overruning NR_IRQS structures in setup_IO_APIC_irqs.
>>>      
>>>
>>  Do you mean the panic on running out of sources should be put
>>back in?
>>    
>>
>
>No, something like this, although i think Linus hates something about it. 
>Note i forgot to subtract 0x80 from the total vector count in the email 
>(should be NR_IRQ_VECTORS=189).
>
>Yes i have a system which oopses due to the high irq count with mainline. 
>The NR_IRQS overrun is one, the vector collisions is another.
>
>Date: Mon, 7 Apr 2003 22:04:13 -0400 (EDT)
>From: Zwane Mwaikambo <zwane@linuxpower.ca>
>X-X-Sender: zwane@montezuma.mastecende.com
>To: Linus Torvalds <torvalds@transmeta.com>
>cc: Linux Kernel <linux-kernel@vger.kernel.org>
>Subject: Re: [PATCH][2.5] avoid scribbling in IDT with high interrupt count.
>In-Reply-To: <Pine.LNX.4.44.0304070818340.26364-100000@home.transmeta.com>
>Message-ID: <Pine.LNX.4.50.0304071958360.21025-100000@montezuma.mastecende.com>
>References: <Pine.LNX.4.44.0304070818340.26364-100000@home.transmeta.com>
>MIME-Version: 1.0
>Content-Type: TEXT/PLAIN; charset=US-ASCII
>
>On Mon, 7 Apr 2003, Linus Torvalds wrote:
>
>  
>
>>Zwane - is there any reason we couldn't just start re-using irq vector
>>offsets when this happens? We already re-use the vectors themselves, so 
>>restarting the offset pointer shouldn't really _change_ anything.
>>    
>>
>
>My patch skips irq numbers > NR_IRQS and simply return error when we run 
>out of vectors, although mainline doesn't assign duplicates and cause 
>collisions, it does waste vector space. e.g. for a system with 
>NR_IRQS = 224 we have 23 vectors free, 0 collisions and 167 useable irqs 
>and assign_irq_vectors states that we're out of vectors.
>
>  
>
>>In other words, I'm wondering if this simpler patch wouldn't be sufficient 
>>instead?
>>
>>Can you please test this, and re-submit (and if you can explain why your 
>>patch is better, please do so - I have nothing fundamentally against it, I 
>>just want to understand _why_ the complexity is needed).
>>    
>>
>
>Your patch booted the system but there are vector collisions resulting in 
>lost irq routing when we program the IOAPIC with the duplicated vector. So 
>with your patch and NR_IRQS = 224 we have 1 vectors free (0x80), 34 collisions 
>and 225 irqs. However that isn't a fault of your patch but a fault with 
>the NR_IRQS definition. This is what the vector space looks like on i386 
>at present.
>
>0________0x31__________________________0xef____________0xff
>  system	   io interrupts            resvd vectors
>
>0xef - 0x31 = 190 useable io interrupt vectors
>
>So perhaps we should, apply your patch, add a NR_IRQ_VECTORS define 
>and also add commentary in irq_vectors.h how does the following look? I 
>had to readd the NR_IRQS checks to protect against overrunning NR_IRQS sized 
>arrays and i added nr_assigned to track how many vectors were allocated so 
>taht we can bail out when we're out.
>
>Patch has been tested on a 320 interrupt system and had a maximum useable 
>irq line of 211 (ethernet).
>
>Thanks,
>	Zwane
>
>Index: linux-2.5.67/include/asm-i386/mach-default/irq_vectors.h
>===================================================================
>RCS file: /build/cvsroot/linux-2.5.67/include/asm-i386/mach-default/irq_vectors.h,v
>retrieving revision 1.1.1.1
>diff -u -p -B -r1.1.1.1 irq_vectors.h
>--- linux-2.5.67/include/asm-i386/mach-default/irq_vectors.h	8 Apr 2003 01:15:29 -0000	1.1.1.1
>+++ linux-2.5.67/include/asm-i386/mach-default/irq_vectors.h	8 Apr 2003 01:34:54 -0000
>@@ -68,15 +68,22 @@
> #define TIMER_IRQ 0
> 
> /*
>- * 16 8259A IRQ's, 208 potential APIC interrupt sources.
>+ * 16 8259A IRQ's, MAX_IRQ_SOURCES-16 potential APIC interrupt sources.
>  * Right now the APIC is mostly only used for SMP.
>  * 256 vectors is an architectural limit. (we can have
>  * more than 256 devices theoretically, but they will
>  * have to use shared interrupts)
>  * Since vectors 0x00-0x1f are used/reserved for the CPU,
>- * the usable vector space is 0x20-0xff (224 vectors)
>+ * the usable vector space is 0x20-0xff (224 vectors).
>+ * Linux currently makes 189 vectors available for io interrupts
>+ * starting at FIRST_DEVICE_VECTOR till FIRST_SYSTEM_VECTOR
>+ * 
>+ * 0________0x31__________________________0xef______0xff
>+ *   system           io interrupts           resvd
>+ *
>  */
> #ifdef CONFIG_X86_IO_APIC
>+#define NR_IRQ_VECTORS	189
> #define NR_IRQS 224
> #else
> #define NR_IRQS 16
>Index: linux-2.5.67/arch/i386/kernel/io_apic.c
>===================================================================
>RCS file: /build/cvsroot/linux-2.5.67/arch/i386/kernel/io_apic.c,v
>retrieving revision 1.1.1.1
>diff -u -p -B -r1.1.1.1 io_apic.c
>--- linux-2.5.67/arch/i386/kernel/io_apic.c	8 Apr 2003 01:15:35 -0000	1.1.1.1
>+++ linux-2.5.67/arch/i386/kernel/io_apic.c	8 Apr 2003 01:36:06 -0000
>@@ -1107,9 +1107,15 @@ int irq_vector[NR_IRQS] = { FIRST_DEVICE
> 
> static int __init assign_irq_vector(int irq)
> {
>-	static int current_vector = FIRST_DEVICE_VECTOR, offset = 0;
>+	static int current_vector = FIRST_DEVICE_VECTOR, offset = 0,
>+		nr_assigned = 1;
>+
> 	if (IO_APIC_VECTOR(irq) > 0)
> 		return IO_APIC_VECTOR(irq);
>+
>+	if (++nr_assigned > NR_IRQ_VECTORS)
>+		return -ENOSPC;
>+
> next:
> 	current_vector += 8;
> 	if (current_vector == SYSCALL_VECTOR)
>@@ -1167,6 +1173,8 @@ void __init setup_IO_APIC_irqs(void)
> 		}
> 
> 		irq = pin_2_irq(idx, apic, pin);
>+		if (irq >= NR_IRQS)
>+			continue;
> 		/*
> 		 * skip adding the timer int on secondary nodes, which causes
> 		 * a small but painful rift in the time-space continuum
>@@ -1181,6 +1189,9 @@ void __init setup_IO_APIC_irqs(void)
> 
> 		if (IO_APIC_IRQ(irq)) {
> 			vector = assign_irq_vector(irq);
>+			if (vector < 0)
>+				continue;
>+
> 			entry.vector = vector;
> 
> 			if (IO_APIC_irq_trigger(irq))
>@@ -2277,6 +2288,10 @@ int io_apic_set_pci_routing (int ioapic,
> {
> 	struct IO_APIC_route_entry entry;
> 	unsigned long flags;
>+	int vector;
>+
>+	if (irq >= NR_IRQS)
>+		return -ENOSPC;
> 
> 	if (!IO_APIC_IRQ(irq)) {
> 		printk(KERN_ERR "IOAPIC[%d]: Invalid reference to IRQ 0/n", 
>@@ -2301,8 +2316,11 @@ int io_apic_set_pci_routing (int ioapic,
> 
> 	add_pin_to_irq(irq, ioapic, pin);
> 
>-	entry.vector = assign_irq_vector(irq);
>+	vector = assign_irq_vector(irq);
>+	if (vector < 0)
>+		return -ENOSPC;
> 
>+	entry.vector = vector;
> 	printk(KERN_DEBUG "IOAPIC[%d]: Set PCI routing entry (%d-%d -> 0x%x -> "
> 		"IRQ %d)\n", ioapic, 
> 		mp_ioapics[ioapic].mpc_apicid, pin, entry.vector, irq);
>
>  
>

Re: [PATCH] 2.5.68 Fix IO_APIC IRQ assignment bug

[Zwane Mwaikambo]

On Mon, 21 Apr 2003, Mika Penttilä wrote:

> Why can't we use the same vector for multiple ioapic entrys? After all, 
> we are already sharing irqs, and an irq is just a cookie for a vector. 
> What do you mean with "lost irq routing" ?

Each ioredtbl can take a vector, if you assign another ioredtbl with the 
same vector and different IRQ then you collide with the previous entry and 
wipe it from the IDT. Also irq != vector

	Zwane

Re: [PATCH] 2.5.68 Fix IO_APIC IRQ assignment bug

[Mika Penttilä]

yes the current code has the assumption of 1to1 mapping from vector to 
irq, but that's a software limitation.

--Mika


Zwane Mwaikambo wrote:

>On Mon, 21 Apr 2003, Mika Penttilä wrote:
>
>  
>
>>Why can't we use the same vector for multiple ioapic entrys? After all, 
>>we are already sharing irqs, and an irq is just a cookie for a vector. 
>>What do you mean with "lost irq routing" ?
>>    
>>
>
>Each ioredtbl can take a vector, if you assign another ioredtbl with the 
>same vector and different IRQ then you collide with the previous entry and 
>wipe it from the IDT. Also irq != vector
>
>	Zwane
>
>-
>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] 2.5.68 Fix IO_APIC IRQ assignment bug

[Maciej W. Rozycki]

On Mon, 21 Apr 2003, [ISO-8859-1] Mika Penttilä wrote:

> Why can't we use the same vector for multiple ioapic entrys? After all, 
> we are already sharing irqs, and an irq is just a cookie for a vector. 

 IIRC, there are serious issues with using the same vector for multiple
I/O APIC interrupts, at least for certain implementations.  So it's
probably not even worth investigating. 

-- 
+  Maciej W. Rozycki, Technical University of Gdansk, Poland   +
+--------------------------------------------------------------+
+        e-mail: macro@ds2.pg.gda.pl, PGP key available        +

Re: [PATCH] 2.5.68 Fix IO_APIC IRQ assignment bug

[Maciej W. Rozycki]

On Mon, 21 Apr 2003, Chuck Ebbert wrote:

>  But doesn't IRQ 0x80, even though it is software-initiated, contend
> with 'real' device interrupts at priority 8, which would mean there are
> three possible sources (80, 81 and 89?)  That's what I was assuming...

 Problems are with local APIC hardware (with queueing arriving IRQ
messages); "int 0x80" doesn't go through the APIC. 

-- 
+  Maciej W. Rozycki, Technical University of Gdansk, Poland   +
+--------------------------------------------------------------+
+        e-mail: macro@ds2.pg.gda.pl, PGP key available        +

Re: [PATCH] 2.5.68 Fix IO_APIC IRQ assignment bug

[Chuck Ebbert]

 Oops, meant to send this to l-k:

Linus Torvalds wrote:

>> Looks like the fix for the "ran out of interrupt sources" panic
>>has a problem.  It will eventually assign a device the same IRQ
>>number as the first system vector, i.e. the local APIC timer.
 ...
> Did you actually see this on hardware?

  In my dreams. :)  But someone must have such hardware or the panic
wouldn't have been removed...

  Only reason I found it at all is I had changed the exact same lines
in my patch (and mine had a much bigger bug than that.)

>Btw, why would you _want_ your redirect table to look like that?
>
>> NR Log Phy Mask Trig IRR Pol Stat Dest Deli Vect:   
>> 00 001 01  0    0    0   0   0    1    1    E7     <== timer at level E
>> 01 001 01  0    0    0   0   0    1    1    30     <== start at 30, not
31

 The patch does two things:

  1.  Reserves all of priority level E for the timers --
      legacy timer at E7, first system vector DF.
      This is might be worth doing (but first system
      vector should be E0.)

  2.  Starts assigning devices at 0x30 instead of 0x31.

(1) is probably worth doing (first system vector should be E0, though.)

> Starting at 31 is better, because..
 ...
> Then we'd have devices at 81 and 89 (two per block of 16 is ok, but 80
> isn't ok because we use that for system calls).
>
> And having two devices in the 8x series means that it takes more irq
> sources to overflow and start to re-use the 3x block - and we want to

 But doesn't IRQ 0x80, even though it is software-initiated, contend
with 'real' device interrupts at priority 8, which would mean there are
three possible sources (80, 81 and 89?)  That's what I was assuming...



------
 Chuck



------
 Chuck

RE: [PATCH] 2.5.68 Fix IO_APIC IRQ assignment bug

[Nakajima, Jun]

> I know there are more people who want to get rid of NR_IRQS e.g. due to
> very sparse irq distribution. For one of the platforms i'm interested in,
> we have to make a clear distinction between irqs and vectors so that we
> can have seperate vector allocations per interrupt handling domain. I
> believe IA-64 does the same but instead per cpu (our domain/node consists
> of 4 cpus) NR_IRQS gets in the way due to it being set at 224 when we
> actually

I heard such requests too, and suggested the same thing, i.e. seperate vector allocations per interrupt handling domain. 

> can service NR_IRQ_VECTORS * NUM_MAXNODES I/O vectors. Can you post your
> patch?

Yes, we'll post it after cleanups.

> 
> Also what MSI devices are you using?
> 

Adaptec 39320 SCSI HBA (it has the 7902 ASIC), which has two MSI-capable functions. You don't need to change the driver (aic79xx) to enable MSI.

Thanks,
Jun

RE: [PATCH] 2.5.68 Fix IO_APIC IRQ assignment bug

[Zwane Mwaikambo]

On Mon, 21 Apr 2003, Nakajima, Jun wrote:

> BTW, I think a better way is to switch from IRQ-base to vector-base. 
> We are working on PCI MSI (Messaged Signaled Support) support, and MSI 
> does not require IRQ (or the platform does not provide IRQ at all), and 
> found a vector-based solution was simpler. 
> 
> Even if MSI is not required, I think vector-based is a clearer solution 
> (IA-64 is using vector-based). IRQs are given by the platform, and the 
> kernel cannot do anything with those. Vector assignment/allocation are 
> fully controlled by the kernel, and the kernel can return the vector 
> number instead of IRQ (except legacy drivers where IRQs < 16). 
> 
> We made a prototype that simply returns the vector numbers for IRQ to 
> device drivers (dev.irq). The function do_IRQ(), for example, gets the 
> vector number, instead of IRQ. No changes to arch/i386/kernel/irq.c were 
> required.

I know there are more people who want to get rid of NR_IRQS e.g. due to 
very sparse irq distribution. For one of the platforms i'm interested in, 
we have to make a clear distinction between irqs and vectors so that we 
can have seperate vector allocations per interrupt handling domain. I 
believe IA-64 does the same but instead per cpu (our domain/node consists 
of 4 cpus) NR_IRQS gets in the way due to it being set at 224 when we actually 
can service NR_IRQ_VECTORS * NUM_MAXNODES I/O vectors. Can you post your 
patch?

Also what MSI devices are you using?

> Before (IRQ-based)
> # cat /proc/interrupts 
>            CPU0       CPU1       
>   0:      10921     671640    IO-APIC-edge  timer
>   2:          0          0          XT-PIC  cascade
>   9:          0          0   IO-APIC-level  acpi
>  14:       5102          1    IO-APIC-edge  ide0
>  15:         10          1    IO-APIC-edge  ide1
>  16:          0          0   IO-APIC-level  uhci-hcd, uhci-hcd
>  18:        449          0   IO-APIC-level  uhci-hcd
>  19:         61          0   IO-APIC-level  uhci-hcd
>  20:        345          0   IO-APIC-level  eth0
>  23:          0          0   IO-APIC-level  ehci-hcd
> NMI:          0          0 
> LOC:     680526     680437 
> ERR:          0
> MIS:          0
> 
> After (vector-based)
>            CPU0       CPU1       
>   0:     709682          0    IO-APIC-edge  timer
>   2:          0          0          XT-PIC  cascade
>   9:          0          0   IO-APIC-level  acpi
>  14:       4988          1    IO-APIC-edge  ide0
>  15:         10          1    IO-APIC-edge  ide1
> 177:         78          0   IO-APIC-level  uhci-hcd
> 185:          0          0   IO-APIC-level  uhci-hcd, uhci-hcd
> 193:         58          0   IO-APIC-level  uhci-hcd
> 201:          0          0   IO-APIC-level  ehci-hcd
> 209:        356          0   IO-APIC-level  eth0
> NMI:          0          0 
> LOC:     707613     707524 
> ERR:          0
> MIS:          0

-- 
function.linuxpower.ca

RE: [PATCH] 2.5.68 Fix IO_APIC IRQ assignment bug

[Nakajima, Jun]

BTW, I think a better way is to switch from IRQ-base to vector-base. 
We are working on PCI MSI (Messaged Signaled Support) support, and MSI does not require IRQ (or the platform does not provide IRQ at all), and found a vector-based solution was simpler. 

Even if MSI is not required, I think vector-based is a clearer solution (IA-64 is using vector-based). IRQs are given by the platform, and the kernel cannot do anything with those. Vector assignment/allocation are fully controlled by the kernel, and the kernel can return the vector number instead of IRQ (except legacy drivers where IRQs < 16). 

We made a prototype that simply returns the vector numbers for IRQ to device drivers (dev.irq). The function do_IRQ(), for example, gets the vector number, instead of IRQ. No changes to arch/i386/kernel/irq.c were required.

Before (IRQ-based)
# cat /proc/interrupts 
           CPU0       CPU1       
  0:      10921     671640    IO-APIC-edge  timer
  2:          0          0          XT-PIC  cascade
  9:          0          0   IO-APIC-level  acpi
 14:       5102          1    IO-APIC-edge  ide0
 15:         10          1    IO-APIC-edge  ide1
 16:          0          0   IO-APIC-level  uhci-hcd, uhci-hcd
 18:        449          0   IO-APIC-level  uhci-hcd
 19:         61          0   IO-APIC-level  uhci-hcd
 20:        345          0   IO-APIC-level  eth0
 23:          0          0   IO-APIC-level  ehci-hcd
NMI:          0          0 
LOC:     680526     680437 
ERR:          0
MIS:          0

After (vector-based)
           CPU0       CPU1       
  0:     709682          0    IO-APIC-edge  timer
  2:          0          0          XT-PIC  cascade
  9:          0          0   IO-APIC-level  acpi
 14:       4988          1    IO-APIC-edge  ide0
 15:         10          1    IO-APIC-edge  ide1
177:         78          0   IO-APIC-level  uhci-hcd
185:          0          0   IO-APIC-level  uhci-hcd, uhci-hcd
193:         58          0   IO-APIC-level  uhci-hcd
201:          0          0   IO-APIC-level  ehci-hcd
209:        356          0   IO-APIC-level  eth0
NMI:          0          0 
LOC:     707613     707524 
ERR:          0
MIS:          0



> -----Original Message-----
> From: Zwane Mwaikambo [mailto:zwane@linuxpower.ca]
> Sent: Monday, April 21, 2003 8:17 AM
> To: Chuck Ebbert
> Cc: linux-kernel
> Subject: Re: [PATCH] 2.5.68 Fix IO_APIC IRQ assignment bug
> 
> On Mon, 21 Apr 2003, Chuck Ebbert wrote:
> 
> > > Yes, we need to bail out in assign_irq_vector when we wrap around,
> > > otherwise we cause collisions when programming the IOAPIC. And we also
> > > need to avoid overruning NR_IRQS structures in setup_IO_APIC_irqs.
> >
> >
> >   Do you mean the panic on running out of sources should be put
> > back in?
> 
> No, something like this, although i think Linus hates something about it.
> Note i forgot to subtract 0x80 from the total vector count in the email
> (should be NR_IRQ_VECTORS=189).
> 
> Yes i have a system which oopses due to the high irq count with mainline.
> The NR_IRQS overrun is one, the vector collisions is another.
> 
> Date: Mon, 7 Apr 2003 22:04:13 -0400 (EDT)
> From: Zwane Mwaikambo <zwane@linuxpower.ca>
> X-X-Sender: zwane@montezuma.mastecende.com
> To: Linus Torvalds <torvalds@transmeta.com>
> cc: Linux Kernel <linux-kernel@vger.kernel.org>
> Subject: Re: [PATCH][2.5] avoid scribbling in IDT with high interrupt
> count.
> In-Reply-To: <Pine.LNX.4.44.0304070818340.26364-100000@home.transmeta.com>
> Message-ID: <Pine.LNX.4.50.0304071958360.21025-
> 100000@montezuma.mastecende.com>
> References: <Pine.LNX.4.44.0304070818340.26364-100000@home.transmeta.com>
> MIME-Version: 1.0
> Content-Type: TEXT/PLAIN; charset=US-ASCII
> 
> On Mon, 7 Apr 2003, Linus Torvalds wrote:
> 
> > Zwane - is there any reason we couldn't just start re-using irq vector
> > offsets when this happens? We already re-use the vectors themselves, so
> > restarting the offset pointer shouldn't really _change_ anything.
> 
> My patch skips irq numbers > NR_IRQS and simply return error when we run
> out of vectors, although mainline doesn't assign duplicates and cause
> collisions, it does waste vector space. e.g. for a system with
> NR_IRQS = 224 we have 23 vectors free, 0 collisions and 167 useable irqs
> and assign_irq_vectors states that we're out of vectors.
> 
> > In other words, I'm wondering if this simpler patch wouldn't be
> sufficient
> > instead?
> >
> > Can you please test this, and re-submit (and if you can explain why your
> > patch is better, please do so - I have nothing fundamentally against it,
> I
> > just want to understand _why_ the complexity is needed).
> 
> Your patch booted the system but there are vector collisions resulting in
> lost irq routing when we program the IOAPIC with the duplicated vector. So
> with your patch and NR_IRQS = 224 we have 1 vectors free (0x80), 34
> collisions
> and 225 irqs. However that isn't a fault of your patch but a fault with
> the NR_IRQS definition. This is what the vector space looks like on i386
> at present.
> 
> 0________0x31__________________________0xef____________0xff
>   system	   io interrupts            resvd vectors
> 
> 0xef - 0x31 = 190 useable io interrupt vectors
> 
> So perhaps we should, apply your patch, add a NR_IRQ_VECTORS define
> and also add commentary in irq_vectors.h how does the following look? I
> had to readd the NR_IRQS checks to protect against overrunning NR_IRQS
> sized
> arrays and i added nr_assigned to track how many vectors were allocated so
> taht we can bail out when we're out.
> 
> Patch has been tested on a 320 interrupt system and had a maximum useable
> irq line of 211 (ethernet).
> 
> Thanks,
> 	Zwane
> 
> Index: linux-2.5.67/include/asm-i386/mach-default/irq_vectors.h
> ===================================================================
> RCS file: /build/cvsroot/linux-2.5.67/include/asm-i386/mach-
> default/irq_vectors.h,v
> retrieving revision 1.1.1.1
> diff -u -p -B -r1.1.1.1 irq_vectors.h
> --- linux-2.5.67/include/asm-i386/mach-default/irq_vectors.h	8 Apr
> 2003 01:15:29 -0000	1.1.1.1
> +++ linux-2.5.67/include/asm-i386/mach-default/irq_vectors.h	8 Apr
> 2003 01:34:54 -0000
> @@ -68,15 +68,22 @@
>  #define TIMER_IRQ 0
> 
>  /*
> - * 16 8259A IRQ's, 208 potential APIC interrupt sources.
> + * 16 8259A IRQ's, MAX_IRQ_SOURCES-16 potential APIC interrupt sources.
>   * Right now the APIC is mostly only used for SMP.
>   * 256 vectors is an architectural limit. (we can have
>   * more than 256 devices theoretically, but they will
>   * have to use shared interrupts)
>   * Since vectors 0x00-0x1f are used/reserved for the CPU,
> - * the usable vector space is 0x20-0xff (224 vectors)
> + * the usable vector space is 0x20-0xff (224 vectors).
> + * Linux currently makes 189 vectors available for io interrupts
> + * starting at FIRST_DEVICE_VECTOR till FIRST_SYSTEM_VECTOR
> + *
> + * 0________0x31__________________________0xef______0xff
> + *   system           io interrupts           resvd
> + *
>   */
>  #ifdef CONFIG_X86_IO_APIC
> +#define NR_IRQ_VECTORS	189
>  #define NR_IRQS 224
>  #else
>  #define NR_IRQS 16
> Index: linux-2.5.67/arch/i386/kernel/io_apic.c
> ===================================================================
> RCS file: /build/cvsroot/linux-2.5.67/arch/i386/kernel/io_apic.c,v
> retrieving revision 1.1.1.1
> diff -u -p -B -r1.1.1.1 io_apic.c
> --- linux-2.5.67/arch/i386/kernel/io_apic.c	8 Apr 2003 01:15:35 -0000
> 	1.1.1.1
> +++ linux-2.5.67/arch/i386/kernel/io_apic.c	8 Apr 2003 01:36:06 -0000
> @@ -1107,9 +1107,15 @@ int irq_vector[NR_IRQS] = { FIRST_DEVICE
> 
>  static int __init assign_irq_vector(int irq)
>  {
> -	static int current_vector = FIRST_DEVICE_VECTOR, offset = 0;
> +	static int current_vector = FIRST_DEVICE_VECTOR, offset = 0,
> +		nr_assigned = 1;
> +
>  	if (IO_APIC_VECTOR(irq) > 0)
>  		return IO_APIC_VECTOR(irq);
> +
> +	if (++nr_assigned > NR_IRQ_VECTORS)
> +		return -ENOSPC;
> +
>  next:
>  	current_vector += 8;
>  	if (current_vector == SYSCALL_VECTOR)
> @@ -1167,6 +1173,8 @@ void __init setup_IO_APIC_irqs(void)
>  		}
> 
>  		irq = pin_2_irq(idx, apic, pin);
> +		if (irq >= NR_IRQS)
> +			continue;
>  		/*
>  		 * skip adding the timer int on secondary nodes, which causes
>  		 * a small but painful rift in the time-space continuum
> @@ -1181,6 +1189,9 @@ void __init setup_IO_APIC_irqs(void)
> 
>  		if (IO_APIC_IRQ(irq)) {
>  			vector = assign_irq_vector(irq);
> +			if (vector < 0)
> +				continue;
> +
>  			entry.vector = vector;
> 
>  			if (IO_APIC_irq_trigger(irq))
> @@ -2277,6 +2288,10 @@ int io_apic_set_pci_routing (int ioapic,
>  {
>  	struct IO_APIC_route_entry entry;
>  	unsigned long flags;
> +	int vector;
> +
> +	if (irq >= NR_IRQS)
> +		return -ENOSPC;
> 
>  	if (!IO_APIC_IRQ(irq)) {
>  		printk(KERN_ERR "IOAPIC[%d]: Invalid reference to IRQ 0/n",
> @@ -2301,8 +2316,11 @@ int io_apic_set_pci_routing (int ioapic,
> 
>  	add_pin_to_irq(irq, ioapic, pin);
> 
> -	entry.vector = assign_irq_vector(irq);
> +	vector = assign_irq_vector(irq);
> +	if (vector < 0)
> +		return -ENOSPC;
> 
> +	entry.vector = vector;
>  	printk(KERN_DEBUG "IOAPIC[%d]: Set PCI routing entry (%d-%d -> 0x%x
> -> "
>  		"IRQ %d)\n", ioapic,
>  		mp_ioapics[ioapic].mpc_apicid, pin, entry.vector, irq);
> 
> --
> function.linuxpower.ca
> -
> 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] 2.5.68 Fix IO_APIC IRQ assignment bug

[Linus Torvalds]

In article <200304201811_MC3-1-3537-1648@compuserve.com>,
Chuck Ebbert  <76306.1226@compuserve.com> wrote:
>
> I found this while trying to forward-port my .66 patch to make
>the redirect table look like this:

Btw, why would you _want_ your redirect table to look like that?

> NR Log Phy Mask Trig IRR Pol Stat Dest Deli Vect:   
> 00 001 01  0    0    0   0   0    1    1    E7     <== timer at level E
> 01 001 01  0    0    0   0   0    1    1    30     <== start at 30, not 31

Starting at 31 is better, because..

> 02 000 00  1    0    0   0   0    0    0    00
> 03 001 01  0    0    0   0   0    1    1    38
> 04 001 01  0    0    0   0   0    1    1    40
> 05 001 01  0    0    0   0   0    1    1    48
> 06 001 01  0    0    0   0   0    1    1    50
> 07 001 01  0    0    0   0   0    1    1    58
> 08 001 01  0    0    0   0   0    1    1    60
> 09 001 01  0    0    0   0   0    1    1    68
> 0a 001 01  0    0    0   0   0    1    1    70
> 0b 001 01  0    0    0   0   0    1    1    78
> 0c 001 01  0    0    0   0   0    1    1    88     <== only one device at 8

Then we'd have devices at 81 and 89 (two per block of 16 is ok, but 80
isn't ok because we use that for system calls).

And having two devices in the 8x series means that it takes more irq
sources to overflow and start to re-use the 3x block - and we want to
delay re-using the 3x block as long as possible due to the silly "only
one pending irq per block" rule that some intel APIC's have.

So that's why FIRST_DEVICE_VECTOR is normally 31 - because it gives a
nicer pattern for the first round of allocations, and that's the common
case (machines with hundreds of APIC irq sources are still rare, the
common case is a single IOAPIC with 24 or so pins).

		Linus

Re: [PATCH] 2.5.68 Fix IO_APIC IRQ assignment bug

[Zwane Mwaikambo]

On Sun, 20 Apr 2003, Linus Torvalds wrote:

> Good call.
> 
> Although I suspect you need about a million interrupt sources to hit
> this, since FIRST_SYSTEM_VECTOR is somethign like 0xef, and thus you can
> hit it only when "offset" has already been incremented seven times
> (which implies that we've walked the whole vector space quite a few
> times by then).
> 
> Did you actually see this on hardware?

Yes, we need to bail out in assign_irq_vector when we wrap around, 
otherwise we cause collisions when programming the IOAPIC. And we also 
need to avoid overruning NR_IRQS structures in setup_IO_APIC_irqs.

	Zwane

Re: [PATCH] 2.5.68 Fix IO_APIC IRQ assignment bug

[Linus Torvalds]

In article <200304201811_MC3-1-3537-1648@compuserve.com>,
Chuck Ebbert  <76306.1226@compuserve.com> wrote:
>
> Looks like the fix for the "ran out of interrupt sources" panic
>has a problem.  It will eventually assign a device the same IRQ
>number as the first system vector, i.e. the local APIC timer.

Good call.

Although I suspect you need about a million interrupt sources to hit
this, since FIRST_SYSTEM_VECTOR is somethign like 0xef, and thus you can
hit it only when "offset" has already been incremented seven times
(which implies that we've walked the whole vector space quite a few
times by then).

Did you actually see this on hardware?

Anyway, applied as obvious.

		Linus

[PATCH] 2.5.68 Fix IO_APIC IRQ assignment bug

[Chuck Ebbert]

 Looks like the fix for the "ran out of interrupt sources" panic
has a problem.  It will eventually assign a device the same IRQ
number as the first system vector, i.e. the local APIC timer.
I think this will fix it:


--- a/arch/i386/kernel/io_apic.c
+++ b/arch/i386/kernel/io_apic.c
@ -1117,7 +1117,7 @
 	if (current_vector == SYSCALL_VECTOR)
 		goto next;
 
-	if (current_vector > FIRST_SYSTEM_VECTOR) {
+	if (current_vector >= FIRST_SYSTEM_VECTOR) {
 		offset = (offset + 1) & 7;
 		current_vector = FIRST_DEVICE_VECTOR + offset;
 	}


 I found this while trying to forward-port my .66 patch to make
the redirect table look like this:


 NR Log Phy Mask Trig IRR Pol Stat Dest Deli Vect:   
 00 001 01  0    0    0   0   0    1    1    E7     <== timer at level E
 01 001 01  0    0    0   0   0    1    1    30     <== start at 30, not 31
 02 000 00  1    0    0   0   0    0    0    00
 03 001 01  0    0    0   0   0    1    1    38
 04 001 01  0    0    0   0   0    1    1    40
 05 001 01  0    0    0   0   0    1    1    48
 06 001 01  0    0    0   0   0    1    1    50
 07 001 01  0    0    0   0   0    1    1    58
 08 001 01  0    0    0   0   0    1    1    60
 09 001 01  0    0    0   0   0    1    1    68
 0a 001 01  0    0    0   0   0    1    1    70
 0b 001 01  0    0    0   0   0    1    1    78
 0c 001 01  0    0    0   0   0    1    1    88     <== only one device at 8
 0d 001 01  0    0    0   0   0    1    1    90
 0e 001 01  0    0    0   0   0    1    1    98
 0f 000 00  1    0    0   0   0    0    0    00
 10 001 01  1    1    0   1   0    1    1    A0
 11 001 01  1    1    0   1   0    1    1    A8
 12 001 01  1    1    0   1   0    1    1    B0
 13 001 01  1    1    0   1   0    1    1    B8
 14 001 01  0    0    0   0   0    1    1    C0
 15 000 00  1    0    0   0   0    0    0    00
 16 000 00  1    0    0   0   0    0    0    00
 17 000 00  1    0    0   0   0    0    0    00



------
 Chuck