[OpenBIOS] Bug in ESP DMA implementation?

[Blue Swirl]

On Fri, Dec 17, 2010 at 6:14 PM, Mark Cave-Ayland
<mark.cave-ayland at siriusit.co.uk> wrote:
> Hi all,
>
> Continuing more with my work on migrating SPARC32 to OFMEM, I've hit an
> issue with the ESP driver which is causing me a bit of a problem.
>
> At the moment, I have a hybrid old-mem/OFMEM SPARC32 setup where I am
> migrating the various memory calls over to OFMEM one at a time. Currently my
> implementation just uses OFMEM for allocating MMU page tables, and with the
> default compile option of -Os looks like this:
>
> Configuration device id QEMU version 1 machine id 32
> Unhandled Exception 0x0000001f
> PC = 0xffd12f08 NPC = 0xffd12f0c
> Stopping execution
>
> The interesting part is that this problem goes away if I compile with any
> -O3 or -O0 or but shows when I compile with -O2, -O1 or -Os. So I wonder if
> I've hit some kind of logic bug in OpenBIOS?
>
> Looking at the SPARCv8 specification, exception 0x1f is equivalent to IRQ15
> and the offending code where the error occurs can be found in
> drivers/esp.c:do_command():
>
>
>    esp->ll->regs[ESP_BUSID] = sd->id & 7;
>    // Set DMA address
>    esp->espdma.regs->st_addr = esp->buffer_dvma;
>    // Set DMA length
>    esp->ll->regs[ESP_TCLOW] = cmdlen & 0xff;
>    esp->ll->regs[ESP_TCMED] = (cmdlen >> 8) & 0xff;
>    // Set DMA direction and enable DMA
>    esp->espdma.regs->cond_reg = DMA_ENABLE;
>
> /* Crash occurs somewhere in this section... */
>
>    // Set ATN, issue command
>    esp->ll->regs[ESP_CMD] = ESP_CMD_SELA | ESP_CMD_DMA;
>    // Wait for DMA to complete. Can this fail?
>    while ((esp->espdma.regs->cond_reg & DMA_HNDL_INTR) == 0)
>
> /* End of crash section */;
>
>    // Check status
>    status = esp->ll->regs[ESP_STATUS];
>    // Clear interrupts to avoid guests seeing spurious interrupts
>    (void)esp->ll->regs[ESP_INTRPT];
>
>
> I notice from the code above there is an explicit comment that mentions
> clearing interrupts to prevent the guest from seeing them so I would have
> thought that this wouldn't be an issue? I've checked the espdma structures
> to ensure that they are marked volatile (or _volatile_) and this appears to
> be the case - so I'm a little bit stumped. Can anyone point me in the right
> direction or spot the mistake?

One tricky case was that when allocating memory for IOMMU, the
alignment restrictions concern physical memory, not virtual.