Re: [PATCH 2/3] powerpc/powernv: DMA operations for discontiguous allocation

[Russell Currey <ruscur@russell.cc>]

[-- Attachment #1: Type: text/plain, Size: 1121 bytes --]

On Fri, 2018-06-29 at 17:34 +1000, Russell Currey wrote:

<snip>

> +		/*
> +		 * The TCE isn't being used, so let's try and
> allocate it.
> +		 * Bits 0 and 1 are read/write, and we use bit 2 as
> a "lock"
> +		 * bit.  This is to prevent any race where the value
> is set in
> +		 * the TCE table but the invalidate/mb() hasn't
> finished yet.
> +		 */
> +		entry = cpu_to_be64((addr - offset) | 7);
> +		ret = cmpxchg(&pe->tces[i], tce, entry);
> +		if (ret != tce) {
> +			/* conflict, start looking again just in
> case */
> +			i--;
> +			continue;
> +		}
> +		pnv_pci_phb3_tce_invalidate(pe, 0, 0, addr - offset,
> 1);

This is wrong and won't work outside of PHB3, will make a generic
handler

> +		mb();
> +		/* clear the lock bit now that we know it's active
> */
> +		ret = cmpxchg(&pe->tces[i], entry, cpu_to_be64((addr
> - offset) | 3));
> +		if (ret != entry) {
> +			/* conflict, start looking again just in
> case */
> +			i--;
> +			continue;
> +		}
> +
> +		return (i << phb->ioda.max_tce_order) | offset;
> +	}
> +	/* If we get here, the table must be full, so error out. */
> +	return -1ULL;
> +}
> +

[-- Attachment #2: qtpass.desktop --]
[-- Type: application/x-desktop, Size: 911 bytes --]