On Monday 25 July 2005 19:38, Martin Egholm Nielsen wrote:
> >>The situation: The mount times of my 32 megs JFFS2 device suddenly
> >>increased from seconds to (roughly) 9 minutes, after having written some
> >>files to the device.
> >>The explanation of this is, according to David Woodhouse's best guess,
> >>that the garbage collector uses all this time "for building up the
> >>node tree for every inode after mounting".
> >>The problem stems (I've been told) from the fact that I have been
> >>performing "big-file-gymnastics" (11 megs uncompressed - ~3 megs
> >>compressed) on the device. Possibly along with some small-file actions
> >>inbetween (?)...
> >>Now my question: Can YAFFS (1&2?) be provoked into showing similar
> >>"unfortunate" behaviour, or is it handled in another way?
> >
> > When I originally evaluated JFFS2, before proposing YAFFS, I identified a
> > few areas of concern wrt running JFFS2 on NAND. One of those was how its
> > garbage collection strategy would scale to large files and large NAND
> > partitions - which seems to be the problem you are running into here.
>
> Exactly... However with the recent CVS version, these 9 minutes reduced
> to ~45sec - hence "only" twice of what it takes to read the entire
> flash-device (raw with dd)...
>
> Will YAFFS' worst case scenario be somewhat like the time it takes to
> perform "dd if=/dev/mtd0 of=/dev/null"? (In my sitation ~20 secs)
There are some differences between YAFFS1 and YAFFS2 here, so I will split
them apart.
To see this in code form, read yaffs_Scan for YAFFS1 scanning and
yaffs_ScanBackwards for YAFFS2 scanning.
The main difference is that YAFFS1 has deleted tags markers while YAFFS2 does
not. This makes the scanning different.
YAFFS1 scanning looks more or less like:
for(all blocks)
for(all written chunks in block)
read tags (ie read oob/spare)
if(!tags.deleted)
{
if(tags says it is an object header)
read whole chunk to extract file info
else
if is a data chunk, insert into tree
}
YAFFS2 scanning looks like:
for(all written blocks backwards)
for(all written chunks in block backwards
read tags (ie read oob/spare)
if(tags says it is an object header and we don't yet have file info)
read whole chunk to extract file info
else if it is a data chunk
if is a data chunk, insert into tree
}
As you can see, in both cases, YAFFS only reads the nand once [well yaffs2
also reads one chunk per block to determine if the block is written first].
YAFFS only makes one pass.
Most chunks will be data chunks and only the oob/spare needs to be read.
The absolutely worst case would be a file system that is full of file headers
(ie. thousands of zero length files). In this case, the whole nand would have
to be read **once**.
>
> > YAFFS does not use compression and has a very clean and simple overwrite
> > and garbage collection model. This makes YAFFS garbage collection and
> > scanning a lot more predictable and cheaper.
>
> I actually could live without the compression - it's the ECC and
> robustness at powerfailure I'm concerned with...
>
> > During mount, both YAFFS and JFFS2 rebuild trees by scanning. YAFFS
> > "cheats" by using the spare/oob area as a place to store tags and by
> > using fixed size chunks. This makes YAFFS scanning pretty fast. (Or
> > course it could be made faster still by saving the state as big binary
> > blobs).
>
> Nice! Is YAFFS2 faster at this?
As always, it depends...
>
> > Now clearly a full YAFFS system will take longer to mount than an empty
> > one, but I don't think it will every get anywhere near as nasty as the
> > case you mention above.
>
> Hopefully not :-)
>
> > BTW: I'm not knocking JFFS2 here, I think it definitely has its place
> > where space is very limited. The transition point is probably around
> > 16MB, depending of course on application needs.
>
> I'm really considering it...
I cannot force you, but I think you will be happy. Most people that have moved
have reported a significant improvement is performance.
-- Charles
(text/plain)