[U-Boot] [PATCH 2/2] fs/fat/fat: Optimizes memory size with single global variable instead of 3

[Benoît Thébaudeau]

Dear Tien Fong,

On Thu, Jul 14, 2016 at 12:48 PM, Tien Fong Chee <tfchee at altera.com> wrote:
> Dear Benoît,
>
> On Wed, 2016-07-13 at 12:56 +0200, Benoît Thébaudeau wrote:
>> Dear Tien Fong Chee,
>>
>> On Jul 13, 2016 at 11:01 AM, Tien Fong Chee wrote:
>> > Single 64KB get_contents_vfatname_block global variable would be
>> > used for
>> > all FAT implementation instead of allocating additional two global
>> > variables
>> > which are get_denfromdir_block and do_fat_read_at_block. This
>> > implementation
>> > can help in saving up 128KB memory space.
>> >
>> > Signed-off-by: Tien Fong Chee <tfchee at altera.com>
>> > Cc: Dinh Nguyen <dinguyen at opensource.altera.com>
>> > Cc: Dinh Nguyen <dinh.linux at gmail.com>
>> > Cc: ChinLiang <clsee at altera.com>
>> > Cc: Vagrant Cascadian <vagrant at debian.org>
>> > Cc: Simon Glass <sjg at chromium.org>
>> > Cc: Stephen Warren <swarren at nvidia.com>
>> > Cc: Benoît Thébaudeau <benoit at wsystem.com>
>> > ---
>> >  fs/fat/fat.c |    6 ++----
>> >  1 files changed, 2 insertions(+), 4 deletions(-)
>> >
>> > diff --git a/fs/fat/fat.c b/fs/fat/fat.c
>> > index 826bd85..5d1afe6 100644
>> > --- a/fs/fat/fat.c
>> > +++ b/fs/fat/fat.c
>> > @@ -579,8 +579,7 @@ static __u8 mkcksum(const char name[8], const
>> > char ext[3])
>> >   * Get the directory entry associated with 'filename' from the
>> > directory
>> >   * starting at 'startsect'
>> >   */
>> > -__u8 get_dentfromdir_block[MAX_CLUSTSIZE]
>> > -   __aligned(ARCH_DMA_MINALIGN);
>> > +__u8 *get_dentfromdir_block = get_contents_vfatname_block;
>> >
>> >  static dir_entry *get_dentfromdir(fsdata *mydata, int startsect,
>> >                               char *filename, dir_entry
>> > *retdent,
>> > @@ -811,8 +810,7 @@ exit:
>> >     return ret;
>> >  }
>> >
>> > -__u8 do_fat_read_at_block[MAX_CLUSTSIZE]
>> > -   __aligned(ARCH_DMA_MINALIGN);
>> > +__u8 *do_fat_read_at_block = get_contents_vfatname_block;
>> >
>> >  int do_fat_read_at(const char *filename, loff_t pos, void *buffer,
>> >                loff_t maxsize, int dols, int dogetsize, loff_t
>> > *size)
>>
>> This probably breaks at least fat_write.c, which uses:
>>   memcpy(get_dentfromdir_block, get_contents_vfatname_block,
>
> With this patch, this line code "memcpy(get_dentfromdir_block,
> get_contents_vfatname_block," is not required anymore since both
>  get_dentfromdir_block and get_contents_vfatname_block are sharing the
> same content and memory. So, this line code can be removed or leaving
> in there. However, there is only one place within fill_dir_slot_buffer
> function where it can corrupt the the global memory, and it is fixed by
> replacing with local buffer. This was sent out with another patch
> called "[PATCH 1/2] fs/fat/fatwrite: Local variable as buffer to store
> dir_slot entries". Overall, applying these two patches, a lot memory
> space can be saved and fitting into small OCRAM, but need to be very
> careful when modifying the code related to global memory.

I get the point, but I am a bit concerned because these changes make
the code even more fragile and hard to maintain than it currently is.
Perhaps it would be time to switch to FatFs as previously suggested.
Here is its memory usage:
http://elm-chan.org/fsw/ff/en/appnote.html#memory

I've not checked in detail all the possible code paths, but for
instance, if I look at get_vfatname(), it's called twice in fat.c,
once with cluster and retdent pointing to somewhere in
get_dentfromdir_block, and once with them pointing to somewhere in
do_fat_read_at_block (through other pointer variables), while
get_vfatname() may write to get_contents_vfatname_block.
get_vfatname() then uses the original data pointed to by slotptr >=
retdent. To make things worse, there is a memcpy (not memmove) from
realdent, which may point to somewhere in get_contents_vfatname_block,
to retdent. It's almost impossible for the involved buffer areas not
to overlap in that case since a whole cluster is read.

I also agree with Stephen's recommendations.

Best regards,
Benoît