Re: [PATCH] BISECTED x86: avoid qword access in memcpy_*io

["H. Peter Anvin" <hpa@zytor.com>]

On 07/20/2010 06:21 PM, Hidetoshi Seto wrote:
> With v2.6.35-rc5, my x86-64 server doesn't boot but reports a
> Completer Abort on lpfc card.
> 
> The result of git-bisect is:
>   6175ddf06b6172046a329e3abfd9c901a43efd2e is the first bad commit
>   commit 6175ddf06b6172046a329e3abfd9c901a43efd2e
>   Author: Brian Gerst <brgerst@gmail.com>
>   Date:   Fri Feb 5 09:37:07 2010 -0500
>     x86: Clean up mem*io functions.
> 
> What I found are:
>  - memcpy for 64bit uses movq if count >= 64 (arch/x86/lib/memcpy_64.S)
>  - memcpy_toio and memcpy_fromio have changed to use this memcpy by
>    the above commit.
>  - my debug shows that lpfc calls memcpy_toio with not-qword-aligned
>    addresses and count >= 64, e.g.:
>      memcpy_toio(0xffffc900118de004, 0xffff88047293d614, 124);
>    and it seems that it comes from:
>    [drivers/scsi/lpfc/lpfc_sli.c]
>     4929   /* First copy mbox command data to HBA SLIM, skip past first
>     4930      word */
>     4931   to_slim = phba->MBslimaddr + sizeof (uint32_t);
>     4932   lpfc_memcpy_to_slim(to_slim, &mb->un.varWords[0],
>     4933               MAILBOX_CMD_SIZE - sizeof (uint32_t));
> 
> Still I'm not sure what is wrong in software or hardware, however
> I suppose that qword access to iomem is not always safe, so it will
> be OK to back to use __inline_memcpy which uses movsl.
> 
> I confirmed that my server (w/ lpfc) boots with 35-rc5 + this patch.
> 
> Signed-off-by: Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>

A driver should not use the memcpy-like instructions if it isn't set up
to act as memory (meaning it can handle arbitrary byte enables.)

The function it should be using is called, fairly counterintuitively,
__iowrite32_copy().  It really should be called memcpy_toio32() or
something similar.

	-hpa

-- 
H. Peter Anvin, Intel Open Source Technology Center
I work for Intel.  I don't speak on their behalf.