[openib-general] Ordering between PCI config space writes and MMIO reads?

[Michael S. Tsirkin]

Quoting r. Roland Dreier <rdreier at cisco.com>:
> Subject: Re: Ordering between PCI config space writes and MMIO reads?
> 
> The discussion fizzled out without really reaching a definitive
> answer, so I'm going to apply the original patch (below), since I
> pretty much convinced myself that only the driver doing the config
> access has enough information to fix this reliably.
> 
>  - R.
> 
> Author: John Partridge <johnip at sgi.com>
> Date:   Tue Oct 31 11:00:04 2006 -0800
> 
>     IB/mthca: Make sure all PCI config writes reach device before doing MMIO
>     
>     During initialization, mthca writes some PCI config space registers
>     and then does an MMIO read from one of the BARs it just enabled.  This
>     MMIO read sometimes failed and caused a crash on SGI Altix machines,
>     because the PCI-X host bridge (legitimately, according to the PCI
>     spec) allowed the MMIO read to start before the config write completed.
>     
>     To fix this, add a config read after all config writes to make sure
>     they are all done before starting the MMIO read.
>     
>     Signed-off-by: John Partridge <johnip at sgi.com>
>     Signed-off-by: Roland Dreier <rolandd at cisco.com>
> 
> diff --git a/drivers/infiniband/hw/mthca/mthca_reset.c b/drivers/infiniband/hw/mthca/mthca_reset.c
> index 91934f2..578dc7c 100644
> --- a/drivers/infiniband/hw/mthca/mthca_reset.c
> +++ b/drivers/infiniband/hw/mthca/mthca_reset.c
> @@ -281,6 +281,20 @@ good:
>  		goto out;
>  	}
>  
> +	/*
> +	 * Perform a "flush" of the PCI config writes here by reading
> +	 * the PCI_COMMAND register.  This is needed to make sure that
> +	 * we don't try to touch other PCI BARs before the config
> +	 * writes are done -- otherwise an MMIO cycle could start
> +	 * before the config writes are done and reach the HCA before
> +	 * the BAR is actually enabled.
> +	 */
> +	if (pci_read_config_dword(mdev->pdev, PCI_COMMAND, hca_header)) {
> +		err = -ENODEV;
> +		mthca_err(mdev, "Couldn't access HCA memory after restoring, "
> +			  "aborting.\n");
> +	}
> +
>  out:
>  	if (bridge)
>  		pci_dev_put(bridge);

Here's what I don't understand: according to PCI rules, pci config read
can bypass pci config write (both are non-posted).
So why does doing it help flush the writes as the comment claims?

Isn't this more the case of
/* pci_config_write seems to complete asynchronously on Altix systems.
 * This is probably broken but its not clear what's the best
 * thing to do is - for now, do pci_read_config_dword which seems to flush
 * everything out. */

-- 
MST