Hi, James,
I think I found the root cause: lpfc's device memory doesn't support 128bit load/store. But Loongson-3's optimised memcpy() will use 128bit load/store instructions (gslq/gssq). So is this patch neccesory to avoid 128bit memory access (maybe there are other platforms have the same problem)?
Huacai
------------------ Original ------------------
From: "James Smart"<james.smart@xxxxxxxxxxxx>;
Date: Mon, Mar 5, 2018 11:37 PM
To: "陈华才"<chenhc@xxxxxxxxxx>; "Dick Kennedy"<dick.kennedy@xxxxxxxxxxxx>;
Cc: "James E . J . Bottomley"<jejb@xxxxxxxxxxxxxxxxxx>; "Martin K . Petersen"<martin.petersen@xxxxxxxxxx>; "Fuxin Zhang"<zhangfx@xxxxxxxxxx>; "linux-scsi"<linux-scsi@xxxxxxxxxxxxxxx>; "stable"<stable@xxxxxxxxxxxxxxx>;
Subject: Re: [PATCH] scsi: lpfc: Switch memcpy_fromio() to __read32_copy()
On 3/4/2018 10:56 PM, 陈华才 wrote:
> No one has the same problem? Or this patch need improve?
>
> Huacai
>
> ------------------ Original ------------------
> From: "陈华才"<chenhc@xxxxxxxxxx>;
> Date: Thu, Mar 1, 2018 01:59 PM
> To: "James Smart"<james.smart@xxxxxxxxxxxx>; "Dick Kennedy"<dick.kennedy@xxxxxxxxxxxx>;
> Cc: "James E . J . Bottomley"<jejb@xxxxxxxxxxxxxxxxxx>; "Martin K . Petersen"<martin.petersen@xxxxxxxxxx>; "Fuxin Zhang"<zhangfx@xxxxxxxxxx>; "linux-scsi"<linux-scsi@xxxxxxxxxxxxxxx>; "stable"<stable@xxxxxxxxxxxxxxx>;
> Subject: Re:[PATCH] scsi: lpfc: Switch memcpy_fromio() to __read32_copy()
>
> Ping?
>
> ------------------ Original ------------------
> From: "Huacai Chen"<chenhc@xxxxxxxxxx>;
> Date: Fri, Jan 26, 2018 08:53 PM
> To: "James Smart"<james.smart@xxxxxxxxxxxx>; "Dick Kennedy"<dick.kennedy@xxxxxxxxxxxx>;
> Cc: "James E . J . Bottomley"<jejb@xxxxxxxxxxxxxxxxxx>; "Martin K . Petersen"<martin.petersen@xxxxxxxxxx>; "Fuxin Zhang"<zhangfx@xxxxxxxxxx>; "linux-scsi"<linux-scsi@xxxxxxxxxxxxxxx>; "Huacai Chen"<chenhc@xxxxxxxxxx>; "stable"<stable@xxxxxxxxxxxxxxx>;
> Subject: [PATCH] scsi: lpfc: Switch memcpy_fromio() to __read32_copy()
>
> In commit bc73905abf770192 ("[SCSI] lpfc 8.3.16: SLI Additions, updates,
> and code cleanup"), lpfc_memcpy_to_slim() have switched memcpy_toio() to
> __write32_copy() in order to prevent unaligned 64 bit copy. Recently, we
> found that lpfc_memcpy_from_slim() have similar issues, so let it switch
> memcpy_fromio() to __read32_copy().
>
> Cc: stable@xxxxxxxxxxxxxxx
> Signed-off-by: Huacai Chen <chenhc@xxxxxxxxxx>
> ---
> drivers/scsi/lpfc/lpfc_compat.h | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/scsi/lpfc/lpfc_compat.h b/drivers/scsi/lpfc/lpfc_compat.h
> index 6b32b0a..47d4fad 100644
> --- a/drivers/scsi/lpfc/lpfc_compat.h
> +++ b/drivers/scsi/lpfc/lpfc_compat.h
> @@ -91,8 +91,8 @@ lpfc_memcpy_to_slim( void __iomem *dest, void *src, unsigned int bytes)
> static inline void
> lpfc_memcpy_from_slim( void *dest, void __iomem *src, unsigned int bytes)
> {
> - /* actually returns 1 byte past dest */
> - memcpy_fromio( dest, src, bytes);
> + /* convert bytes in argument list to word count for copy function */
> + __ioread32_copy(dest, src, bytes / sizeof(uint32_t));
> }
>
> #endif /* __BIG_ENDIAN */
Huacai,
Note: the original change was 7 yrs ago and we've had no reports of
issues. but....
This change didn't look necessary as the memcpy_XXXio() should allow
byte alignments and byte-based lengths. Which means the original patch
also wasn't necessary, which implies that the platform that had the
issues has an api restriction not normally found on the other platforms.
I have no problem with the change though.
-- james
Signed-off-by: James Smart <james.smart@xxxxxxxxxxxx>
