[gentoo-embedded] file system question

[gentoo-embedded] file system question

[David Relson]

G'day,

I'm porting the software for an embedded medical device from DOS to
Linux and am wondering which file systems are appropriate and which are
not.  The device's mass storage is a Disk-on-Module solid state flash
drive.  Data is presently written at approx 100 bytes every 30 seconds
but that might change to 100 bytes every second.  The device has a
watchdog (recently activated) and during today's session it was
triggered and wiped out my file system.

Anybody have recommendations on which file system to use and the
appropriate settings?

Anybody have suggested readings so I can educate myself?

Thank you.

David

Re: [gentoo-embedded] file system question

[Ingo Krabbe]

On Mon, Mar 29, 2010 at 06:42:15PM -0400, David Relson wrote:
> G'day,
> 
> I'm porting the software for an embedded medical device from DOS to
> Linux and am wondering which file systems are appropriate and which are
> not.  The device's mass storage is a Disk-on-Module solid state flash
> drive.  Data is presently written at approx 100 bytes every 30 seconds
> but that might change to 100 bytes every second.  The device has a
> watchdog (recently activated) and during today's session it was
> triggered and wiped out my file system.
> 
> Anybody have recommendations on which file system to use and the
> appropriate settings?
> 
> Anybody have suggested readings so I can educate myself?

First I would read about Wear Levelling (Wikipedia).  Maybe your device already
implements TrueFFS or ExtremeFFS with low-level wear levelling, so it might be
enough to just use any other file system upon.  Then I would choose a simple
ext2 file system, though I can't tell the wear levelling really works.  Are
there methods to debug that?

I think the best choose for you might be a JFFS2 Filesystem.  Or just choose one
from "Flash file system" in Wikipedia.

bye ingo

Re: [gentoo-embedded] file system question

[David Relson]

On Tue, 30 Mar 2010 10:40:25 +0200
Ingo Krabbe wrote:

> On Mon, Mar 29, 2010 at 06:42:15PM -0400, David Relson wrote:
> > G'day,
> > 
> > I'm porting the software for an embedded medical device from DOS to
> > Linux and am wondering which file systems are appropriate and which
> > are not.  The device's mass storage is a Disk-on-Module solid state
> > flash drive.  Data is presently written at approx 100 bytes every
> > 30 seconds but that might change to 100 bytes every second.  The
> > device has a watchdog (recently activated) and during today's
> > session it was triggered and wiped out my file system.
> > 
> > Anybody have recommendations on which file system to use and the
> > appropriate settings?
> > 
> > Anybody have suggested readings so I can educate myself?
> 
> First I would read about Wear Levelling (Wikipedia).  Maybe your
> device already implements TrueFFS or ExtremeFFS with low-level wear
> levelling, so it might be enough to just use any other file system
> upon.  Then I would choose a simple ext2 file system, though I can't
> tell the wear levelling really works.  Are there methods to debug
> that?
> 
> I think the best choose for you might be a JFFS2 Filesystem.  Or just
> choose one from "Flash file system" in Wikipedia.
> 
> bye ingo

Hello Ingo,

I've been using ext2 and have encountered some problems.  In addition
to porting from DOS to Linux, the PC-104 board that has been used for
years has reached end-of-life, so there's a new PC-104 board that's
also involved.  With the new board life has been good.  With the old
board a file system problem has been encountered many times.  

For reasons totally unclear (at least to me), all of a sudden a file
can't be read or a shared library doesn't load.  When I check using ls, 
the message "Stale NFS file handle" appears.  The device has _never_
been on a network, so NFS  doesn't actually play any role in the
device's life cycle.

This has been happening at least a couple of times a week for the last
month.  Often it happens several times in the same day.  The file
handle problem has only happened with the old board, never with the
new.  As there are several thousand "old board" devices in use, the
Linux port needs to support them.  

FWIW, the cpu is a 486 and we're using the 2.6.29.6 kernel.

Regards,

David

P.S.  The "flash file system" suggestion is appreciated!  I'll check
into it.

Re: [gentoo-embedded] file system question

[Manuel Lauss]

On Tue, Mar 30, 2010 at 1:47 PM, David Relson <relson@osagesoftware.com> wrote:
> For reasons totally unclear (at least to me), all of a sudden a file
> can't be read or a shared library doesn't load.  When I check using ls,
> the message "Stale NFS file handle" appears.  The device has _never_

Ext2 throws these when it's damaged.  Happens to me usually on unclean
or missing
unmounts (i.e. yanking power cord).

Maybe mounting it with "sync" may alleviate this somewhat on the
problematic boards.

Manuel

Re: [gentoo-embedded] file system question

[Karl Hiramoto]

On 03/30/2010 10:40 AM, Ingo Krabbe wrote:
> On Mon, Mar 29, 2010 at 06:42:15PM -0400, David Relson wrote:
>    
>> G'day,
>>
>> I'm porting the software for an embedded medical device from DOS to
>> Linux and am wondering which file systems are appropriate and which are
>> not.  The device's mass storage is a Disk-on-Module solid state flash
>> drive.  Data is presently written at approx 100 bytes every 30 seconds
>> but that might change to 100 bytes every second.  The device has a
>> watchdog (recently activated) and during today's session it was
>> triggered and wiped out my file system.
>>
>> Anybody have recommendations on which file system to use and the
>> appropriate settings?
>>
>> Anybody have suggested readings so I can educate myself?
>>      
> First I would read about Wear Levelling (Wikipedia).  Maybe your device already
> implements TrueFFS or ExtremeFFS with low-level wear levelling, so it might be
> enough to just use any other file system upon.  Then I would choose a simple
> ext2 file system, though I can't tell the wear levelling really works.  Are
> there methods to debug that?
>
> I think the best choose for you might be a JFFS2 Filesystem.  Or just choose one
> from "Flash file system" in Wikipedia.
>
> bye ingo
>    

You should ignore this advice about a wear leveling FS, if your 
Disk-on-module device does it's own internal wear leveling.  Check your 
module specs.   For example Compact Flash modules have their own 
internal wear leveling and a wear leveling FS will do more writes to the 
flash then necessary.

Re: [gentoo-embedded] file system question

[Karl Hiramoto]

On 03/30/2010 12:42 AM, David Relson wrote:
> G'day,
>
> I'm porting the software for an embedded medical device from DOS to
> Linux and am wondering which file systems are appropriate and which are
> not.  The device's mass storage is a Disk-on-Module solid state flash
> drive.  Data is presently written at approx 100 bytes every 30 seconds
> but that might change to 100 bytes every second.  The device has a
> watchdog (recently activated) and during today's session it was
> triggered and wiped out my file system.
>    
It sounds like some kind of data logging application.   If this is the 
case what i would do is.

1.  Mount read only a partition that contains the main system, to ensure 
you can always boot and to avoid damage to the systems file

2. Mount read/write the partition that data gets written to.  On boot 
you can detect errors in this partition and reformat it if is totally 
hosed.

3.  In your application call fsync() on the file and perhaps the 
directory after the writes to make sure the data gets to the flash.  If 
you are in fact logging, rotate your logs files so after they are X size 
write to a new file.   Files that are most likely to be damaged are ones 
that are opened and being written to the moment the power cord gets yanked.

As Manual said, using the 'sync' option may help.   Also see "man mount" 
and  /usr/src/linux/Documentation/filesystems/ext3.txt

Other options the improve reliability are  data=journal  and barrier=1


I've used ext3 and riserfs on CompactFlash but neither seems to be 100% 
error proof.  I've wanted to try NILFS2 but haven't done it yet.

--
Karl

Re: [gentoo-embedded] file system question

[Peter Stuge]

David Relson wrote:
> The device has a watchdog (recently activated) and during today's
> session it was triggered and wiped out my file system.
> 
> Anybody have recommendations on which file system to use and the
> appropriate settings?

I would consider ext3. In order to deal with sudden resets (or power
loss) you really want a journalling filesystem. In that case you can
lose data (and need to check for, and handle, that on startup) but at
least the filesystem will be structurally ok.

Which filesystem to use on the flash module depends mostly on the
circumstances of systems in the field, such as how often they are
serviced, service personnel skill level and so on.


> Anybody have suggested readings so I can educate myself?

Hm, not so much. :\ The situation you're in is probably pretty
common, I have customers in the exact same situation, but it is very
different from what JFFS2, YAFFS and friends are made for, which is
that the filesystem actually has direct contact with the flash array.

Any kind of consumer level flash module, including disk-on-modules,
CF, SD, you name it, implements wear-levelling transparently. And
everyone considers their particular wear-levelling to be an important
competitive advantage, so they'll never give out details. The likes
of JFFS2, YAFFS require immediate addressing of NAND flash chips
connected to some bus, but that's never the case unless you're making
your own PCBs.


//Peter

RE: [gentoo-embedded] file system question

[Relson, David]

Manuel,

I like the "sync" suggestion and shall add it to fstab shortly.

Regards,
 
David
-----Original Message-----
From: Manuel Lauss [mailto:manuel.lauss@googlemail.com] 
Sent: Tuesday, March 30, 2010 8:18 AM
To: gentoo-embedded@lists.gentoo.org
Cc: ikrabbe.ask@web.de
Subject: Re: [gentoo-embedded] file system question

On Tue, Mar 30, 2010 at 1:47 PM, David Relson <relson@osagesoftware.com> wrote:
> For reasons totally unclear (at least to me), all of a sudden a file
> can't be read or a shared library doesn't load.  When I check using ls,
> the message "Stale NFS file handle" appears.  The device has _never_

Ext2 throws these when it's damaged.  Happens to me usually on unclean
or missing
unmounts (i.e. yanking power cord).

Maybe mounting it with "sync" may alleviate this somewhat on the
problematic boards.

Manuel

Re: [gentoo-embedded] file system question

[Arkadi Shishlov]

On 03/30/10 16:28, Karl Hiramoto wrote:
> I've used ext3 and riserfs on CompactFlash but neither seems to be 100%
> error proof.  I've wanted to try NILFS2 but haven't done it yet.

Let us know then, because something must be done to nilfs2_cleanerd which like
to write a lot.

Re: [gentoo-embedded] file system question

[Peter Stuge]

Karl Hiramoto wrote:
> I've used ext3 and riserfs on CompactFlash

I would avoid reiserfs3. It can very easily become corrupted beyond
repair on power failure.


//Peter

Re: [gentoo-embedded] file system question

[Arkadi Shishlov]

On 03/30/10 16:28, Karl Hiramoto wrote:
> On 03/30/2010 12:42 AM, David Relson wrote:
>> I'm porting the software for an embedded medical device from DOS to
>> Linux and am wondering which file systems are appropriate and which are
>> not.  The device's mass storage is a Disk-on-Module solid state flash
> 
> 1.  Mount read only a partition that contains the main system, to ensure
> you can always boot and to avoid damage to the systems file

Depending on how you want to service (or not) you device root fs, consider using
at SquashFS for that.

Re: [gentoo-embedded] file system question

[Ed W]

On 29/03/2010 23:42, David Relson wrote:
> G'day,
>
> I'm porting the software for an embedded medical device from DOS to
> Linux and am wondering which file systems are appropriate and which are
> not.  The device's mass storage is a Disk-on-Module solid state flash
> drive.  Data is presently written at approx 100 bytes every 30 seconds
> but that might change to 100 bytes every second.  The device has a
> watchdog (recently activated) and during today's session it was
> triggered and wiped out my file system.
>
> Anybody have recommendations on which file system to use and the
> appropriate settings?
>
> Anybody have suggested readings so I can educate myself?
>    

In addition to what everyone else has already noted you really need to 
state your tradeoff between lifetime of the flash cards and minimising 
the risk of loosing data.  Fundamentally you can buffer the data and 
write less frequently to the device (your device has a finite number of 
writes before it perishes), or you can write more frequently, hence less 
risk of dataloss

Any journal'ed fs is *supposed* to be reasonably robust against disk 
crashes during writes, but often there are layered levels of safety, eg 
ext3 has three write modes depending how you want to trade safety for 
speed (hint newest defaults are probably not optimal if your data is 
very precious)

I think one of the issues you will face is that whilst a journaling fs 
with appropriate options should be very safe against you flipping the 
power off whenever you feel like, long term your failure mode with flash 
appears to be random files disappearing and getting corrupted - the wear 
levelling appears to mean that unrelated files can get trashed as the 
flash shuffles data around to maximise wear levels until the card 
basically pretty much gives up?  I haven't experienced this yet...

I guess just plan ahead for this.

Good luck

Ed W

Re: [gentoo-embedded] file system question

[Karl Hiramoto]

On 03/30/2010 04:14 PM, Arkadi Shishlov wrote:
> On 03/30/10 16:28, Karl Hiramoto wrote:
>    
>> I've used ext3 and riserfs on CompactFlash but neither seems to be 100%
>> error proof.  I've wanted to try NILFS2 but haven't done it yet.
>>      
> Let us know then, because something must be done to nilfs2_cleanerd which like
> to write a lot.
>    
Sounds like a bad idea on flash then.

Re: [gentoo-embedded] file system question

[Arkadi Shishlov]

On 03/30/10 18:07, Karl Hiramoto wrote:
> On 03/30/2010 04:14 PM, Arkadi Shishlov wrote:
>> On 03/30/10 16:28, Karl Hiramoto wrote:
>>   
>>> I've used ext3 and riserfs on CompactFlash but neither seems to be 100%
>>> error proof.  I've wanted to try NILFS2 but haven't done it yet.
>>>      
>> Let us know then, because something must be done to nilfs2_cleanerd
>> which like
>> to write a lot.
>>    
> Sounds like a bad idea on flash then.

Its probably depends on workload a lot, plus how often cleanerd is allowed to
work or is it enabled at all.
It actually might be better for flash, cause its not hammering multiple blocks
with journal and metadata (sync) or double writes of data=journal. Working in
sequential manner is presumably advantageous for (cheap) FTL too.
Just a guess... Need someone to try it out in real product. :)

The built-in checkpoint/snapshot feature is nice to have.

Re: [gentoo-embedded] file system question

[wireless]

Ed W wrote:

> In addition to what everyone else has already noted you really need to
> state your tradeoff between lifetime of the flash cards and minimising
> the risk of loosing data.  


Wow, this sounds like a PC programmer's statement and not
how an embedded designer would solve the problem. Granted
the poster has not provided and adequate 'specification' so
my point of view may or may not be precisely applicable.


From what I have gleaned, the developer should first look at
a low power system, with battery backup. If it is an
embedded system designed to run as such, a simple 9V battery
should provide years/decades of reliability to 5 nines of
uptime or more..... This robust research and design approach
will solve power cycling-data loss problems....


Depending on the details of the data collection sensors and
hardware, either redesign as low power, or put them on a
separate power supply. If that, (relatively) hi power supply
is intermittent, you only loose data for those data
collection intervals. You CPU/storage module never looses
power. Some very fast AD devices are power hogs. Other
(slower) AD devices are very power efficient, especially if
they 'sleep' (low or off power mode) when not
actively needed. This is controlled



I could go on and on, but just purchasing a SBC and trying
to turn it into a robust product, is rarely successful in my
experience. USE and SBC of the shelf to get your coders
coding, and get a hardware designer to look at the
specification, and design a low power, minimized design that
has what you need and nothing else.

Then when you ramp up quantity, you'll actually make a
profit as your competition lowers their price, you can still
compete. It's impossible, in my experience, to use off the
shelf SBC and be competitive for very long. Your competition
will under cut you every time....


hth,
James

Re: [gentoo-embedded] file system question

[David Relson]

Hello, All!

The many suggestions have been very helpful.  This afternoon,
'sync' was added to fstab.  No problems have been seen since.
While not conclusive, the indication is good and we're helpful.

The device in question is a medical device that continuously takes
readings, graphs them, and every 30 seconds appends a 34 byte data
record to each of two files.  It's not exactly disk intensive :->

FWIW, often the files hit by the "Stale NFS file handle" problem are
symlinks in /usr/lib.  Since _nobody_ writes that directory it's odd
that the problem often shows up there.  That's not the only place, but
seems to be the most common one.

The suggestion of separate partitions for program and data is one I
like and shall implement.  Thanks for the idea!

The power cord isn't a problem -- there's an onboard battery.  However
the power button can be pressed at any time.  'Tis something that we'll
live with.  Adding fsync() calls, in addition to 'sync' mounting, should
help.

Ciao!

David


On Tue, 30 Mar 2010 15:28:59 +0200
Karl Hiramoto wrote:

> On 03/30/2010 12:42 AM, David Relson wrote:
> > G'day,
> >
> > I'm porting the software for an embedded medical device from DOS to
> > Linux and am wondering which file systems are appropriate and which
> > are not.  The device's mass storage is a Disk-on-Module solid state
> > flash drive.  Data is presently written at approx 100 bytes every
> > 30 seconds but that might change to 100 bytes every second.  The
> > device has a watchdog (recently activated) and during today's
> > session it was triggered and wiped out my file system.
> >    
> It sounds like some kind of data logging application.   If this is
> the case what i would do is.
> 
> 1.  Mount read only a partition that contains the main system, to
> ensure you can always boot and to avoid damage to the systems file
> 
> 2. Mount read/write the partition that data gets written to.  On boot 
> you can detect errors in this partition and reformat it if is totally 
> hosed.
> 
> 3.  In your application call fsync() on the file and perhaps the 
> directory after the writes to make sure the data gets to the flash.
> If you are in fact logging, rotate your logs files so after they are
> X size write to a new file.   Files that are most likely to be
> damaged are ones that are opened and being written to the moment the
> power cord gets yanked.
> 
> As Manual said, using the 'sync' option may help.   Also see "man
> mount" and  /usr/src/linux/Documentation/filesystems/ext3.txt
> 
> Other options the improve reliability are  data=journal  and barrier=1
> 
> 
> I've used ext3 and riserfs on CompactFlash but neither seems to be
> 100% error proof.  I've wanted to try NILFS2 but haven't done it yet.
> 
> --
> Karl
> 
> 
> 

Re: [gentoo-embedded] file system question

[Peter Stuge]

wireless wrote:
> design a low power, minimized design

Sure, except that's not economical below a certain threshold, say 1k
units. Maybe even 10k.

There are lots of great embedded products without mass market, but
they're still important. Making electronics is just too expensive,
and also requires a large range of special knowledge not so easy to
whip out of the sleeve.

I don't contest that custom design will be the best fit in the second
or third hardware generation, but that is quite often traded away for
other benefits - and the products still have to work..

Backup power is fairly easy to accomplish in any case, more of a
money question, but since there's a watchdog the system does need to
handle unclean shutdown in any case.


//Peter

Re: [gentoo-embedded] file system question

[Arkadi Shishlov]

On 03/31/10 03:44, David Relson wrote:
> The many suggestions have been very helpful.  This afternoon,
> 'sync' was added to fstab.  No problems have been seen since.
> While not conclusive, the indication is good and we're helpful.
> 
> The device in question is a medical device that continuously takes
> readings, graphs them, and every 30 seconds appends a 34 byte data
> record to each of two files.  It's not exactly disk intensive :->

Even with data amount is low, you're still looking at something like ~3 x16KB
(eraseblock, which could be significantly larger) number of erases cause some
data should be updated in place. I would be writing into one file with writes
aligned to flash pagesize.
If someone hits the power button while flash is working, sync won't help on
ext2. Backup power is present, why not to wire power button so that shutdown can
be signaled and emergency sync() is performed on filesystem...
While ext3 is mature, ext4 has journal checksums.

> FWIW, often the files hit by the "Stale NFS file handle" problem are
> symlinks in /usr/lib.  Since _nobody_ writes that directory it's odd
> that the problem often shows up there.  That's not the only place, but
> seems to be the most common one.

Weird errors may indicate toolchain problems with in errno.h not matching kernel
version.

> The power cord isn't a problem -- there's an onboard battery.  However
> the power button can be pressed at any time.  'Tis something that we'll
> live with.  Adding fsync() calls, in addition to 'sync' mounting, should
> help.

I believe it is same and you need only one method.

Re: [gentoo-embedded] file system question

[Nebojša Ćosić]

> G'day,
> 
> I'm porting the software for an embedded medical device from DOS to
> Linux and am wondering which file systems are appropriate and which are
> not.  The device's mass storage is a Disk-on-Module solid state flash
> drive.  Data is presently written at approx 100 bytes every 30 seconds
> but that might change to 100 bytes every second.  The device has a
> watchdog (recently activated) and during today's session it was
> triggered and wiped out my file system.
> 
> Anybody have recommendations on which file system to use and the
> appropriate settings?
> 
> Anybody have suggested readings so I can educate myself?
> 
> Thank you.
> 
> David
> 
After having problems with EMC and usb storage, I finally fixed the
problem with following solution:
- data storage, in my case usb stick, has at least 2 partitions
- second partition is without file system. It is divided in a number of
  slots, each large enough to store all of my data
- all work is performed on data stored on ram disk
- periodically (triggered by time and/or data change), I compress ram
  disk and dump it in a next slot on unformatted partition
I have a small battery, which I use to do one final dump at shutdown
time.
On startup, I go through all of the slots in second partition,
searching for latest uncorrupt data, and use this to populate ram disk.
If you can live with some data loss, you don't even need battery backup.
No matter wear leveling implementation on your storage, this solution
works optimally.
It works even on your directly accessible flash storage.
Since there is no real file system on partition, there is no need for
it's recovery - besides searching for latest and greatest set of data
on startup
And it is implemented as a ash script, using tar and gzip, so your data
is actually better verified than on normal file system (the usual one
do not actually checksum data. I don't consider jffs2 to be "the usual
filesystem":)
Nebojša

Re: [gentoo-embedded] file system question

[David Relson]

On Wed, 31 Mar 2010 11:26:49 +0200
Nebojša Ćosić wrote:

> > G'day,
> > 
> > I'm porting the software for an embedded medical device from DOS to
> > Linux and am wondering which file systems are appropriate and which
> > are not.  The device's mass storage is a Disk-on-Module solid state
> > flash drive.  Data is presently written at approx 100 bytes every
> > 30 seconds but that might change to 100 bytes every second.  The
> > device has a watchdog (recently activated) and during today's
> > session it was triggered and wiped out my file system.
> > 
> > Anybody have recommendations on which file system to use and the
> > appropriate settings?
> > 
> > Anybody have suggested readings so I can educate myself?
> > 
> > Thank you.
> > 
> > David
> > 
> After having problems with EMC and usb storage, I finally fixed the
> problem with following solution:
> - data storage, in my case usb stick, has at least 2 partitions
> - second partition is without file system. It is divided in a number
> of slots, each large enough to store all of my data
> - all work is performed on data stored on ram disk
> - periodically (triggered by time and/or data change), I compress ram
>   disk and dump it in a next slot on unformatted partition
> I have a small battery, which I use to do one final dump at shutdown
> time.
> On startup, I go through all of the slots in second partition,
> searching for latest uncorrupt data, and use this to populate ram
> disk. If you can live with some data loss, you don't even need
> battery backup. No matter wear leveling implementation on your
> storage, this solution works optimally.
> It works even on your directly accessible flash storage.
> Since there is no real file system on partition, there is no need for
> it's recovery - besides searching for latest and greatest set of data
> on startup
> And it is implemented as a ash script, using tar and gzip, so your
> data is actually better verified than on normal file system (the
> usual one do not actually checksum data. I don't consider jffs2 to be
> "the usual filesystem":)
> Nebojša

Wow!  That's a robust solution!

Re: [gentoo-embedded] file system question

[Ed W]

On 30/03/2010 19:11, wireless wrote:
> Ed W wrote:
>
>    
>> In addition to what everyone else has already noted you really need to
>> state your tradeoff between lifetime of the flash cards and minimising
>> the risk of loosing data.
>>      
>
> Wow, this sounds like a PC programmer's statement and not
> how an embedded designer would solve the problem. Granted
> the poster has not provided and adequate 'specification' so
> my point of view may or may not be precisely applicable.
>    


...snip...

I don't disagree that you can design the system to loose power or crash 
less frequently, but I don't see how this undermines or disagrees with 
my statement?

Essentially you can buffer more data and hence risk more in the case of 
a crash/power loss, or you can write more frequently which will decrease 
the life of a write limited medium.  I don't see how you have changed 
that basic trade?

The engineering solution is to set things to balance write frequency 
such that the media lasts "long enough", then as you say tackle the 
other side and make the device crash/loose power less frequently.

Based on the nature of the post I assumed that the OP wanted advice on 
tackling the software side...

Regards

Ed W