stm32f4 interrupt pipeline hazard

[mail2]

Posted on October 21, 2013 at 10:26

Hello again,

I have learned that this:

void DMA2_Stream3_IRQHandler() 
{
// Test if DMA Stream Transfer Complete interrupt
if (DMA_GetITStatus(DMA2_Stream3, DMA_IT_TCIF3)) 
{
DMA_ClearITPendingBit(DMA2_Stream3, DMA_IT_TCIF3);
}
}

is danger on a Cortex-M3/4. The Interrupt can be reentered immediately due to some pipline issues. Is it enough to insert two nops to clear this situation?

void DMA2_Stream3_IRQHandler() 
{
// Test if DMA Stream Transfer Complete interrupt
if (DMA_GetITStatus(DMA2_Stream3, DMA_IT_TCIF3)) 
{
DMA_ClearITPendingBit(DMA2_Stream3, DMA_IT_TCIF3);
}
__NOP();
__NOP();
}

Currently it works for me, but does it work under all situations? Martin

[Tesla DeLorean]

Posted on October 21, 2013 at 14:32

What is the point of such a routine? If you do any read/write on memory it will fence the interrupt clearing write. So increment a variable, write a flag, anything should be sufficient.

https://community.st.com/0D50X00009XkYyUSAV

 

https://community.st.com/0D50X00009XkZ4PSAV

 

Edit: Fixed DEAD LINKs, original post from Oct 21, 2013

[mail2]

Posted on October 21, 2013 at 16:39

Hi Clive,

These where dummies, so i can't forget one and i can keep die handler happy while coding and testing. These costs me a few hours of debugging for nothing before i found my fault :)Maybe a big TODO comment, these nops or a __DMB() is better in future (it seems that Keil-C is able to optimize nop() avay on higher optimization settings)Martin

[Tesla DeLorean]

Posted on October 21, 2013 at 16:47

Writing or incrementing a volatile variable should be quite adequate to force in-order completion.

volatile int jiffies;
void DMA2_Stream3_IRQHandler()
{
// Test if DMA Stream Transfer Complete interrupt
if (DMA_GetITStatus(DMA2_Stream3, DMA_IT_TCIF3)) // Qualify
{
DMA_ClearITPendingBit(DMA2_Stream3, DMA_IT_TCIF3); // Clear Early
// Do something
jiffies++;
}
}