[U-Boot] [PATCH v7 2/7] ARM: socfpga: Add default FPGA bitstream fitImage for Arria10 SoCDK

[Marek Vasut]

On 2/11/19 7:23 AM, Chee, Tien Fong wrote:
> On Tue, 2019-02-05 at 09:51 +0100, Marek Vasut wrote:
>> On 2/1/19 5:50 PM, Chee, Tien Fong wrote:
>>>
>>> On Fri, 2019-02-01 at 09:29 +0100, Marek Vasut wrote:
>>>>
>>>> On 2/1/19 4:59 AM, Chee, Tien Fong wrote:
>>>>>
>>>>>
>>>>> On Thu, 2019-01-31 at 15:54 +0100, Marek Vasut wrote:
>>>>>>
>>>>>>
>>>>>> On 1/31/19 3:51 PM, tien.fong.chee at intel.com wrote:
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> From: Tien Fong Chee <tien.fong.chee at intel.com>
>>>>>>>
>>>>>>> Add default fitImage file bundling FPGA bitstreams for
>>>>>>> Arria10.
>>>>>>>
>>>>>>> Signed-off-by: Tien Fong Chee <tien.fong.chee at intel.com>
>>>>>>> ---
>>>>>>>  board/altera/arria10-socdk/fit_spl_fpga.its | 31
>>>>>>> +++++++++++++++++++++++++++++
>>>>>>>  1 file changed, 31 insertions(+)
>>>>>>>  create mode 100644 board/altera/arria10-
>>>>>>> socdk/fit_spl_fpga.its
>>>>>>>
>>>>>>> diff --git a/board/altera/arria10-socdk/fit_spl_fpga.its
>>>>>>> b/board/altera/arria10-socdk/fit_spl_fpga.its
>>>>>>> new file mode 100644
>>>>>>> index 0000000..46b125c
>>>>>>> --- /dev/null
>>>>>>> +++ b/board/altera/arria10-socdk/fit_spl_fpga.its
>>>>>>> @@ -0,0 +1,31 @@
>>>>>>> +// SPDX-License-Identifier: GPL-2.0
>>>>>>> + /*
>>>>>>> + * Copyright (C) 2019 Intel Corporation <www.intel.com>
>>>>>>> + *
>>>>>>> + */
>>>>>>> +
>>>>>>> +/dts-v1/;
>>>>>>> +
>>>>>>> +/ {
>>>>>>> +	description = "FIT image with FPGA bistream";
>>>>>>> +	#address-cells = <1>;
>>>>>>> +
>>>>>>> +	images {
>>>>>>> +		fpga-2 {
>>>>>> Why is fpga-2 before fpga-1 ?
>>>>> 1. The main purpose is for solving the performance issue as i
>>>>> described
>>>>> in cover letter. We can decide the absolute data position for
>>>>> core
>>>>> image, and ensure it is in allignment.
>>>> Where does the alignment problem happen exactly ?
>>> The allignment problem happen in get_contents function, line 373,
>>> at
>>> fs/fat/fat.c .
>> But then you're trying to work around a memcpy performance pentalty
>> in
>> VFAT code by frobbing with file position within a fitImage ? This can
>> not work, since the file alignment within fitImage is not guaranteed
> Yes, setting the absolute data position for the large core rbf file in
> fitImage.
> 
> so, when generating the fitImage through mkimage, you need to set the
> absolute position as argument to -p. Absolute data position is always
> fixed offset based on fitImage base.

This can not work, consider the different filesystems the fitImage can
be stored on. It's not just VFAT with one cluster size, it can be any
configuration of VFAT U-Boot supports, or any filesystem U-Boot
supports. And then the performance penalty could be back.

The proper fix is to optimize what VFAT does, if that is a problem.
Maybe the block cache can help here ?

>>> This happens only when reading offset from a file,
>>> that's why absolute position is very important to set the right
>>> offset
>>> for the core image. The performance penalty can be significantly
>>> incurred with large size core image.
>>>
>>> 		filesize -= actsize;
>>> 		actsize -= pos;
>>> 		memcpy(buffer, tmp_buffer + pos, actsize);
>>> 		free(tmp_buffer);
>>> 		*gotsize += actsize;
>>> 		if (!filesize)
>>> 			return 0;
>>> 		buffer += actsize; <= buffer sometimes is altered to  
>>>                                       unaligned
>>>
>>> This function is basically finding the cluster where the pos
>>> resides
>>> in, adjusting the pos, actsize and file size accordingly when the
>>> base
>>> being changed from beginning of the file to the beginning of the
>>> cluster where the pos resides in.
>>>
>>> Then copying the actsize size of content from pos to the end of the
>>> cluster to the buffer above, and updating buffer to the next write.
>>> The
>>> updated buffer can be unaligned especially the pos is not being set
>>> properly, hence we need the absolute position to fix that.
>>>
>>> When the unaligned buffer is passed as argument to the get_cluster
>>> function, you would see the print out of "FAT: Misaligned buffer
>>> address" at line 264 in that function. A very slow disk_read would
>>> be
>>> implemented to transfer the sector by sector content to the
>>> unaligned
>>> buffer.
>> Can this be fixed then ?
> I have tried few ideas, no one of them work.
> 
> Could you help to take a look?

I am tremendously overloaded, sorry. Try taking a look at the block
cache , drivers/block/blkcache.c , maybe this can help ?

>>
>>>
>>>>
>>>>
>>>> Anyway, you cannot rely on this, the alignment within the
>>>> fitImage
>>>> may
>>>> be changed just by using different strings in the ITS file.
>>> No change for absolute position, it is always same offset based on
>>> the
>>> beginning of a FIT.
>> Try adding a few properties here and there and/or changing the length
>> of
>> some of the strings, you'll see the file offset changes.
> Absolute data position is always fixed offset from base of fitImage
> regardless how many properties are added or changing. But, there is
> potential the overlapping would be happended if data position is too
> close with fit.
>>
>>>
>>>>
>>>>
>>>>>
>>>>>
>>>>> 2. Users know where is the data position for core, so easy for
>>>>> them
>>>>> to
>>>>> program themself with series commands on U-Boot console.
>>>> You should use imxtract to pull out the file from fitImage and
>>>> then
>>>> program it. imxtract can refer to image name, so there's no need
>>>> to
>>>> access raw data within the fitImage by offset.
>>> Yes, that's one of the most effective way. Another is using fatload
>>> with offset.
>> No, it is not, because you do not know the offset. imxtract parses
>> the
>> fitImage structure and computes the offset for you.
> You know the offset, this is absolute offset from the base of fitImage,
> you set it as argument to -p when running mkimage such as "-p 400"
> 
> tools/mkimage -E -p 400 -f board/altera/arria10-socdk/fit_spl_fpga.its
> fit_spl_fpga.itb
You're just working around the alignment problem for one specific
configuration of the VFAT, this does not scale.

-- 
Best regards,
Marek Vasut