Lightweight file server on RPI(NTFS mounts, swap, etc)

kfmut

Hi,

for a long time I was using old RPI2 as a lightweight file server and torrents client but its SD-card with Raspbian has treacherously died the other day so I’m one more time transferring RPI to the NetBSD.

My archive is on an SSD-drive with NTFS filesystem connected via USB and when I try to mount it I get “Device not configured” error. I’ve tried both methods, via kernel builtin module and via FUSE.

mount -t ntfs /dev/sd1a /mnt/
mount_ntfs: /dev/sd1a on /mnt: Device not configured

mount.ntfs-3g /dev/sd1a /mnt/
Error opening '/dev/sd1a': Device not configured
Failed to mount '/dev/sd1a': Device not configured
Either the device is missing or it's powered down, or you have
SoftRAID hardware and must use an activated, different device under
/dev/mapper/, (e.g. /dev/mapper/nvidia_eahaabcc1) to mount NTFS.
Please see the 'dmraid' documentation for help.

fdisk /dev/sd1
…
Partition table:
0: NTFS, OS/2 HPFS, QNX2 or Advanced UNIX (sysid 7)
    start 2048, size 468858880 (228935 MB, Cyls 0-29185/61/60)
1: <UNUSED>
2: <UNUSED>
3: <UNUSED>
No active partition.

Maybe I’m doing something obviously wrong?

Regarding swap file/partition, is it possible to make it on usb-stick? Will it interfere with order in which partitions are mounted and swap added to the system? Is it at least a bearable idea?

rvp

kfmut You may have to use the wedge device (dkNN) instead of sd1a. Look through the output of dkctl sd1 listwedges

kfmut

rvp dkctl replies with "no wedges configured"

rvp

kfmut Try mount /dev/sd1e .... (Use disklabel sd1 to figure out the disk name if you have more than one partition on sd1)

kfmut

rvp Yep, you are right, disklabel shows two partitions and e partition could be mounted, but why fdisk provides false info on disk partitioning is beyond me.

Guess it wouldn’t be possible to mount ssd-drive by it's uuid?

rvp

kfmut but why fdisk provides false info on disk partitioning is beyond me.

It isn't exactly false--just different.

fdisk provides a MS-DOS view of the disk. The disklabel program tells you how NetBSD views the same disk (in the absence of a GPT layout where wedges are created automatically).

Every disk on NetBSD has a disklabel (whether real or fictitious). If there is an actual disklabel, then that will be used. If there is none, then, on MBR partitioned disks, a fictitious label is created by the kernel and the partitions listed in the MBR are assigned starting from letter e. a is "reserved" for a NetBSD root partition; b for swap; c covers the entire NetBSD partition in the MBR. d is the whole disk. (The last two letters are how things are on amd64, and it looks like, arm as well. Other arches might use c for the whole disk.)

kfmut Guess it wouldn’t be possible to mount ssd-drive by it's uuid?

If you use GPT, then you can mount using the GPT label: mount -t type NAME=GPT_label ...

kfmut

rvp Thank you for your clarification!

Mosfet

I would not recommend using a NTFS formated drive for serving files in a lan. NTFS in netbsd runs in userspace (through FUSE) and consumes a considerable amount of CPU resources. In other words, your raspberry will struggle.

In a similar setup I had many years ago ( raspeberry pi 1 era 8-) ), I used a dedicated USB hard drive formated in native netbsd filesystem (ffs). That made a hell of a difference.

kfmut

Mosfet Under Raspbian it wasn't that bad, but I'll check it. Thank you!

Mosfet

kfmut I don't keep track what Raspbian state is, since I had given away my pi a couple of years ago, but if it's linux kernel is >= 5.15 then it has native (kernel) support for NTFS. That is magnitude times faster and lighter, than any userspace (FUSE) alternative.

kfmut

Mosfet Looks like NetBSD currently also has builtin kernel module for ntfs.

kfmut

Guys, could you be so kind and help me with GPT-partitioning or point me to a good tutorial? As a test with mounting partition by GUID via fstab I’ve tried to transfer my USB-stick from MBR to GPT via GParted toolkit and it’s working OK(single fat32 partition) but when I’m doing same thing with gpt tool(per man page) and newfs_msdos(newfs_msdos -F 32 -L KTEST dk0) in NetBSD I’ve unexpected stalls with ls operation, umount, writes via samba from Windows machine, reading is OK as a performance of USB-stick in Windows environment. Only difference I could see between two approaches is that GParted starts partition from sector in 2k range and has some free space after partition, gpt tool by default permits to start it right after primary GPT table and to use all space:

bash-5.1# gpt -r show sd1
     start      size  index  contents
         0         1         PMBR
         1         1         Pri GPT header
         2        32         Pri GPT table
        34  15633341      1  GPT part - Windows basic data
  15633375        32         Sec GPT table
  15633407         1         Sec GPT header

Also partition GUID hasn't changed after I've recreated it with gpt in NetBSD, for some reason I've expected that it would but it's not.

rvp

kfmut in NetBSD I’ve unexpected stalls with ls operation, umount, writes via samba from Windows machine

Writes to MS-DOS filesystems are slower than writes to the native UFS. Or, it could be the USB-stick firmware doing block remapping (USB write speeds can vary quite a bit when this is happening); or, it may simply be a drive slowly going bad...

BTW, what cluster size did newfs_msdos report when you formatted the drive?

kfmut

I've tried to partition the USB-stick with GParted again and got the same weird behavior.

Then with gpt I’ve changed the partition type to ffs and formatted it with newfs -O 2 dk0. At first glance it works as expected. So maybe it’s actually related to filesystem type…for now I’m too tired to dig into this more, maybe I’ll do it tomorrow.

New partition table:

bash-5.1# gpt -r show sd0
     start      size  index  contents
         0         1         PMBR
         1         1         Pri GPT header
         2        32         Pri GPT table
        34      2014         Unused
      2048  15630336      1  GPT part - NetBSD FFSv1/FFSv2
  15632384       991         Unused
  15633375        32         Sec GPT table
  15633407         1         Sec GPT header

kfmut

rvp Writes to MS-DOS filesystems are slower than writes to the native UFS.

It's just stalls completely, command couldn't be interrupted with Ctrl+C, couldn't be killed via kill -TERM <process_id> or even via kill -KILL <process_id>, samba’s share couldn’t be accessed via network during those stalls.

rvp Or, it could be the USB-stick firmware doing block remapping (USB write speeds can vary quite a bit when this is happening)

Today I copied files to a new drive with ffs on it from an old SSD-drive with NTFS filesystem mounted through FUSE ntfs-3g(PERFUSE_BUFSIZE=135168 was set) and experienced preasumbly same behavior, after that I’ve remounted it through built-in NTFS kernel module and everything has copied OK. In the first case cp twice(!!!) has copied only first file and stalled around 14MB mark, with built-in module I’ve copied over 50GB

rvp BTW, what cluster size did newfs_msdos report when you formatted the drive?.

Sorry, I do not remember, it was the default for an 8GB stick, I had only used -F 32 and -L options. If it’s important I could check it.

rvp

kfmut Today I copied files to a new drive with ffs on it from an old SSD-drive with NTFS filesystem mounted through FUSE ntfs-3g(PERFUSE_BUFSIZE=135168 was set) and experienced preasumbly same behavior

Sounds like a PUFFS issue now. (And I'm confused...)

Could you describe exactly how to reproduce the problem?

kfmut

rvp Could you describe exactly how to reproduce the problem?

I was unable to reproduce that with a random set of data but then I remembered that the problem was with a partially downloaded file from transmission torrent-client(those .part files). I think they are not completely allocated on the filesystem, so I’ve tried cp -rv *.part in torrents folder on old SSD-drive and it stuck again. Not sure how to independently reproduce that, but I’m using ARMv7 port( https://cdn.netbsd.org/pub/NetBSD/NetBSD-9.3/evbarm-earmv7hf/binary/gzimg/armv7.img.gz ), source drive was with NTFS filesystem, mounted via PERFUSE_BUFSIZE=135168 mount -t ntfs-3g -o windows_names,uid=32767,gid=32766,ro <device> <mount_point> (uid and gid are for nobody:nogroup), transmission-2.94 client, files are not fully allocated when downloaded via client.

But I guess early problems with msdos filesystem were different in nature, I need to look into that once again...

rvp

kfmut OK. A few points:

Any write operation would've failed on that mount point because of the ro parameter you're passing to ntfs-3g
You're very brave to write a NTFS partition--even with ntfs-3g. (If you want to physically pass data between OSes, I would recommend either VFAT or ext2. If you use SMB or NFS, then the underlying FS doesn't matter)
FUSE, because it operates in userspace can be very slow even for reading.

kfmut

rvp Actually I've changed my plans and used ffs filesystem on new drive, PERFUSE_BUFSIZE for ntfs-3g is not easy to use with fstab and writing with ffs is noticeable faster. On the other hand, the difference between built-in NTFS and FUSE is below 10% for reading via WiFi-network and Samba, numbers were 6MB/s with 90% system idle versus 5.3MB/s and 80-85% system idle.

Small hiccup though with external drives for a RPI, mount in NetBSD is missing nofail option so if those drives are not connected system is unable to load.

JuvenalUrbino

I was thinking of writing a tutorial on NFS on NetBSD (server and client side), to describe my setup. Possibly helpful for your use case.

kfmut

JuvenalUrbino that would be helpful, documentation for NetBSD is a bit scarce in modern ways of doing things: no gpt partitioning, no automount for USB-sticks and SSD-drives, no description of workflow with modern filesystems from Windows and Linux worlds, device naming is confusing with MBR partitioning.