[PATCH v5 01/20] EDAC/synopsys: Fix ECC status data and IRQ disable race condition

[Serge Semin]

On Mon, May 06, 2024 at 12:20:29PM +0200, Borislav Petkov wrote:
> On Thu, Apr 25, 2024 at 03:52:38PM +0300, Serge Semin wrote:
> > Even if we get to add the spin-lock serializing the ECCCLR writes it
> > won't solve the problem since the IRQ-disabler critical section could
> > be executed a bit before the IRQ-handler critical section so the later
> > one will just re-enable the IRQs disabled by the former one.
> > 
> > Here is what is suggested in my patch to fix the problem:
> > 
> >      IRQ-handler                        |    IRQ-disabler
> >                                         |
> > zynqmp_get_error_info:                  |
> >                                         | lock_irqsave
> >                                         | ECCCLR = 0; // disable IRQs
> >                                         | unlock_irqrestore
> >  lock_irqsave;                          |
> >  tmp = ECCCLR | clear_sts_bits;         |
> >  ECCCLR = tmp;                          |
> >  unlock_irqrestore;                     |
> 

> <--- I'm presuming here the IRQ-disabler will reenable interrupts at
> some point?
> 
> Otherwise we have the same problem as before when interrupts remain off
> after the IRQ handler has run.

In the sketch above the IRQ-disabler is the method which disables the
IRQ in the concurrent manner. After my patch is applied the
IRQ-handler will no longer touch the IRQ enable/disable bits, but will
preserve them as is:
-       clearval = ECC_CTRL_CLR_CE_ERR | ECC_CTRL_CLR_CE_ERRCNT;
-       clearval |= ECC_CTRL_CLR_UE_ERR | ECC_CTRL_CLR_UE_ERRCNT;
+       spin_lock_irqsave(&priv->reglock, flags);
+
+       clearval = readl(base + ECC_CLR_OFST) |
+                  ECC_CTRL_CLR_CE_ERR | ECC_CTRL_CLR_CE_ERRCNT |
+                  ECC_CTRL_CLR_UE_ERR | ECC_CTRL_CLR_UE_ERRCNT;
        writel(clearval, base + ECC_CLR_OFST);
-       writel(0x0, base + ECC_CLR_OFST);
+
+       spin_unlock_irqrestore(&priv->reglock, flags);

Thus there won't be need in the IRQs re-enabling later in the handler:

@@ -576,8 +601,6 @@ static irqreturn_t intr_handler(int irq, void *dev_id)
        /* v3.0 of the controller does not have this register */
        if (!(priv->p_data->quirks & DDR_ECC_INTR_SELF_CLEAR))
                writel(regval, priv->baseaddr + DDR_QOS_IRQ_STAT_OFST);
-       else
-               enable_intr(priv);

So the only IRQ-disabler left in the driver - disable_intr() - will be
called from the device/driver remove() function. The ECCCLR CSR access
will be guarded with the spin-lock in the IRQ-disabler and in the
IRQ-handler. So it will be safe to have them executed concurrently.

> 
> Other than that, yes, I see it, we will need the locking.
> 
> Thanks for elaborating.

Always welcome. Glad we've settled this.)

-Serge(y)

> 
> -- 
> Regards/Gruss,
>     Boris.
> 
> https://people.kernel.org/tglx/notes-about-netiquette