Patchwork sparc32 fix spurious dma interrupts

login
register
mail settings
Submitter Artyom Tarasenko
Date Feb. 10, 2010, 10:40 p.m.
Message ID <1265841648-27249-1-git-send-email-atar4qemu@google.com>
Download mbox | patch
Permalink /patch/45066/
State New
Headers show

Comments

Artyom Tarasenko - Feb. 10, 2010, 10:40 p.m.
Don't raise interrupt when not enabled.
Don't set DMA_INTR bit spuriously.
Don't print misleading debug messages "Raise IRQ" when not raising any.

Signed-off-by: Artyom Tarasenko <atar4qemu@gmail.com>
---
Artyom Tarasenko - Feb. 12, 2010, 6:40 p.m.
2010/2/10 Artyom Tarasenko <atar4qemu@googlemail.com>:
> Don't raise interrupt when not enabled.
> Don't set DMA_INTR bit spuriously.

> Don't print misleading debug messages "Raise IRQ" when not raising any.

Haven't noticed that these were introduced recently.
Shall we revert 787cfbc432bf1d353a77cbdb613754f3963371a3 and rebase?

>
> Signed-off-by: Artyom Tarasenko <atar4qemu@gmail.com>
> ---
> diff --git a/hw/sparc32_dma.c b/hw/sparc32_dma.c
> index 6e991e0..b2992ca 100644
> --- a/hw/sparc32_dma.c
> +++ b/hw/sparc32_dma.c
> @@ -125,9 +125,11 @@ static void dma_set_irq(void *opaque, int irq, int level)
>  {
>     DMAState *s = opaque;
>     if (level) {
> -        DPRINTF("Raise IRQ\n");
>         s->dmaregs[0] |= DMA_INTR;
> -        qemu_irq_raise(s->irq);
> +        if (s->dmaregs[0] & DMA_INTREN) {
> +            DPRINTF("Raise IRQ\n");
> +            qemu_irq_raise(s->irq);
> +        }
>     } else {
>         s->dmaregs[0] &= ~DMA_INTR;
>         DPRINTF("Lower IRQ\n");
> @@ -142,8 +145,6 @@ void espdma_memory_read(void *opaque, uint8_t *buf, int len)
>     DPRINTF("DMA read, direction: %c, addr 0x%8.8x\n",
>             s->dmaregs[0] & DMA_WRITE_MEM ? 'w': 'r', s->dmaregs[1]);
>     sparc_iommu_memory_read(s->iommu, s->dmaregs[1], buf, len);
> -    DPRINTF("Raise IRQ\n");
> -    s->dmaregs[0] |= DMA_INTR;
>     s->dmaregs[1] += len;
>  }
>
> @@ -154,8 +155,6 @@ void espdma_memory_write(void *opaque, uint8_t *buf, int len)
>     DPRINTF("DMA write, direction: %c, addr 0x%8.8x\n",
>             s->dmaregs[0] & DMA_WRITE_MEM ? 'w': 'r', s->dmaregs[1]);
>     sparc_iommu_memory_write(s->iommu, s->dmaregs[1], buf, len);
> -    DPRINTF("Raise IRQ\n");
> -    s->dmaregs[0] |= DMA_INTR;
>     s->dmaregs[1] += len;
>  }
>
>
Blue Swirl - Feb. 12, 2010, 7:27 p.m.
On Thu, Feb 11, 2010 at 12:40 AM, Artyom Tarasenko
<atar4qemu@googlemail.com> wrote:
> Don't raise interrupt when not enabled.
> Don't set DMA_INTR bit spuriously.
> Don't print misleading debug messages "Raise IRQ" when not raising any.

This breaks most of my Linux tests. *BSD are unaffected. For example
sparc-test 2.0:

eth0: LANCE 52:54:00:12:34:56
esp0: IRQ 36 SCSI ID 7 Clk 40MHz CCYC=25000 CCF=8 TOut 167 NCR53C9XF(espfast)
ESP: Total of 1 ESP hosts found, 1 actually in use.
scsi0 : Sparc ESP100A-FAST
esp0: Aborting command
esp0: dumping state
esp0: dma -- cond_reg<a4400311> addr<f0010e64>
esp0: SW [sreg<11> sstep<04> ireg<18>]
esp0: HW reread [sreg<93> sstep<00> ireg<10>]
esp0: current command [tgt<02> lun<00> pphase<CLUELESS> cphase<DATAIN>]
esp0: disconnected
esp0: Aborting command
esp0: dumping state
esp0: dma -- cond_reg<a4400310> addr<f0010e64>
esp0: SW [sreg<11> sstep<04> ireg<18>]
esp0: HW reread [sreg<03> sstep<04> ireg<00>]
esp0: current command [tgt<02> lun<00> pphase<UNISSUED> cphase<UNISSUED>]
esp0: disconnected
esp0: Resetting scsi bus
esp0: SCSI bus reset interrupt
esp0: Aborting command
esp0: dumping state
esp0: dma -- cond_reg<a4400211> addr<f000d007>
esp0: SW [sreg<03> sstep<04> ireg<80>]
esp0: HW reread [sreg<91> sstep<04> ireg<18>]
esp0: current command [tgt<02> lun<00> pphase<UNISSUED> cphase<UNISSUED>]
esp0: disconnected
scsi: Device offlined - not ready after error recovery: host 0 channel
0 id 2 lun 0

Since Open/NetBSD still works, there may be yet again another bug that
your patch uncovers.
Artyom Tarasenko - Feb. 12, 2010, 10:32 p.m.
2010/2/12 Blue Swirl <blauwirbel@gmail.com>:
> On Thu, Feb 11, 2010 at 12:40 AM, Artyom Tarasenko
> <atar4qemu@googlemail.com> wrote:
>> Don't raise interrupt when not enabled.
>> Don't set DMA_INTR bit spuriously.
>> Don't print misleading debug messages "Raise IRQ" when not raising any.
>
> This breaks most of my Linux tests. *BSD are unaffected. For example
> sparc-test 2.0:
>
> eth0: LANCE 52:54:00:12:34:56
> esp0: IRQ 36 SCSI ID 7 Clk 40MHz CCYC=25000 CCF=8 TOut 167 NCR53C9XF(espfast)
> ESP: Total of 1 ESP hosts found, 1 actually in use.
> scsi0 : Sparc ESP100A-FAST
> esp0: Aborting command
> esp0: dumping state
> esp0: dma -- cond_reg<a4400311> addr<f0010e64>
> esp0: SW [sreg<11> sstep<04> ireg<18>]
> esp0: HW reread [sreg<93> sstep<00> ireg<10>]
> esp0: current command [tgt<02> lun<00> pphase<CLUELESS> cphase<DATAIN>]
> esp0: disconnected
> esp0: Aborting command
> esp0: dumping state
> esp0: dma -- cond_reg<a4400310> addr<f0010e64>
> esp0: SW [sreg<11> sstep<04> ireg<18>]
> esp0: HW reread [sreg<03> sstep<04> ireg<00>]
> esp0: current command [tgt<02> lun<00> pphase<UNISSUED> cphase<UNISSUED>]
> esp0: disconnected
> esp0: Resetting scsi bus
> esp0: SCSI bus reset interrupt
> esp0: Aborting command
> esp0: dumping state
> esp0: dma -- cond_reg<a4400211> addr<f000d007>
> esp0: SW [sreg<03> sstep<04> ireg<80>]
> esp0: HW reread [sreg<91> sstep<04> ireg<18>]
> esp0: current command [tgt<02> lun<00> pphase<UNISSUED> cphase<UNISSUED>]
> esp0: disconnected
> scsi: Device offlined - not ready after error recovery: host 0 channel
> 0 id 2 lun 0
>
> Since Open/NetBSD still works, there may be yet again another bug that
> your patch uncovers.

Looks like Linux has problems with the fist part of the patch: not
raising irqs when they are not enabled.
The part is actually not relevant for Solaris, so we could just skip
it. But it would be nicer to find this another bug.
It seems that there is some asymmetry: Solaris complained about
spurious interrupts only on write operations.
Do you have an idea why setting the DMA_INTR bit and raising the irq
was split between multiple functions?
Blue Swirl - Feb. 13, 2010, 7:52 a.m.
On Sat, Feb 13, 2010 at 12:32 AM, Artyom Tarasenko
<atar4qemu@googlemail.com> wrote:
> 2010/2/12 Blue Swirl <blauwirbel@gmail.com>:
>> On Thu, Feb 11, 2010 at 12:40 AM, Artyom Tarasenko
>> <atar4qemu@googlemail.com> wrote:
>>> Don't raise interrupt when not enabled.
>>> Don't set DMA_INTR bit spuriously.
>>> Don't print misleading debug messages "Raise IRQ" when not raising any.
>>
>> This breaks most of my Linux tests. *BSD are unaffected. For example
>> sparc-test 2.0:
>>
>> eth0: LANCE 52:54:00:12:34:56
>> esp0: IRQ 36 SCSI ID 7 Clk 40MHz CCYC=25000 CCF=8 TOut 167 NCR53C9XF(espfast)
>> ESP: Total of 1 ESP hosts found, 1 actually in use.
>> scsi0 : Sparc ESP100A-FAST
>> esp0: Aborting command
>> esp0: dumping state
>> esp0: dma -- cond_reg<a4400311> addr<f0010e64>
>> esp0: SW [sreg<11> sstep<04> ireg<18>]
>> esp0: HW reread [sreg<93> sstep<00> ireg<10>]
>> esp0: current command [tgt<02> lun<00> pphase<CLUELESS> cphase<DATAIN>]
>> esp0: disconnected
>> esp0: Aborting command
>> esp0: dumping state
>> esp0: dma -- cond_reg<a4400310> addr<f0010e64>
>> esp0: SW [sreg<11> sstep<04> ireg<18>]
>> esp0: HW reread [sreg<03> sstep<04> ireg<00>]
>> esp0: current command [tgt<02> lun<00> pphase<UNISSUED> cphase<UNISSUED>]
>> esp0: disconnected
>> esp0: Resetting scsi bus
>> esp0: SCSI bus reset interrupt
>> esp0: Aborting command
>> esp0: dumping state
>> esp0: dma -- cond_reg<a4400211> addr<f000d007>
>> esp0: SW [sreg<03> sstep<04> ireg<80>]
>> esp0: HW reread [sreg<91> sstep<04> ireg<18>]
>> esp0: current command [tgt<02> lun<00> pphase<UNISSUED> cphase<UNISSUED>]
>> esp0: disconnected
>> scsi: Device offlined - not ready after error recovery: host 0 channel
>> 0 id 2 lun 0
>>
>> Since Open/NetBSD still works, there may be yet again another bug that
>> your patch uncovers.
>
> Looks like Linux has problems with the fist part of the patch: not
> raising irqs when they are not enabled.
> The part is actually not relevant for Solaris, so we could just skip
> it. But it would be nicer to find this another bug.
> It seems that there is some asymmetry: Solaris complained about
> spurious interrupts only on write operations.
> Do you have an idea why setting the DMA_INTR bit and raising the irq
> was split between multiple functions?

I'd suspect my sloppiness. ;-)

Patch

diff --git a/hw/sparc32_dma.c b/hw/sparc32_dma.c
index 6e991e0..b2992ca 100644
--- a/hw/sparc32_dma.c
+++ b/hw/sparc32_dma.c
@@ -125,9 +125,11 @@  static void dma_set_irq(void *opaque, int irq, int level)
 {
     DMAState *s = opaque;
     if (level) {
-        DPRINTF("Raise IRQ\n");
         s->dmaregs[0] |= DMA_INTR;
-        qemu_irq_raise(s->irq);
+        if (s->dmaregs[0] & DMA_INTREN) {
+            DPRINTF("Raise IRQ\n");
+            qemu_irq_raise(s->irq);
+        }
     } else {
         s->dmaregs[0] &= ~DMA_INTR;
         DPRINTF("Lower IRQ\n");
@@ -142,8 +145,6 @@  void espdma_memory_read(void *opaque, uint8_t *buf, int len)
     DPRINTF("DMA read, direction: %c, addr 0x%8.8x\n",
             s->dmaregs[0] & DMA_WRITE_MEM ? 'w': 'r', s->dmaregs[1]);
     sparc_iommu_memory_read(s->iommu, s->dmaregs[1], buf, len);
-    DPRINTF("Raise IRQ\n");
-    s->dmaregs[0] |= DMA_INTR;
     s->dmaregs[1] += len;
 }
 
@@ -154,8 +155,6 @@  void espdma_memory_write(void *opaque, uint8_t *buf, int len)
     DPRINTF("DMA write, direction: %c, addr 0x%8.8x\n",
             s->dmaregs[0] & DMA_WRITE_MEM ? 'w': 'r', s->dmaregs[1]);
     sparc_iommu_memory_write(s->iommu, s->dmaregs[1], buf, len);
-    DPRINTF("Raise IRQ\n");
-    s->dmaregs[0] |= DMA_INTR;
     s->dmaregs[1] += len;
 }