NetBSD block devices

20-100

I've found no newcomer-friendly documentation about NetBSD's block devices and I'm trying to write one. I've collected information from the various man pages involved but some points remain unclear to me:

Why is a character device created for each block device? The man pages say it is a "raw device", without further explanation.
Why are all possible combinations of unit + partition created under /dev even though there's no corresponding device in the machine?
How can one find which devices are really used and what hardware they represent (besides reading dmesg)?

kamil

20-100

Why is a character device created for each block device? The man pages say it is a "raw device", without further explanation.

A character device is a subset of functionality of a block device, or a block device is a superset of a character device.

Why are all possible combinations of unit + partition created under /dev even though there's no corresponding device in the machine?

NetBSD, by design, does not follow the dynamic devfs model, instead the /dev files/nodes are static and upon open() the kernel can start some devices and load kernel modules behind the scenes. In Linux, FreeBSD, MacOS, DragonFly this is the opposite. You need to load a kernel driver to see a node created in /dev.

How can one find which devices are really used and what hardware they represent (besides reading dmesg)?

drvctl -lt is a good starting point. It depends on the type of a device. Some of them can be probed with sysctl, e.g.:

$ sysctl hw.disknames
hw.disknames = wd0 cgd0

rvp

20-100 Why is a character device created for each block device?

Accessing a raw devices bypasses the disk/buffer cache--which is needed for programs like fsck which scan a lot of the disk and you wouldn't want a cache eviction by its actions. Raw devices also have fairly strict read/write size contraints because they bypass the buffer cache. Eg:

#  dd if=/dev/rsd0 of=/dev/null bs=128 skip=0 count=1
dd: /dev/rsd0: Invalid argument
0+0 records in
0+0 records out
0 bytes transferred in 0.001 secs (0 bytes/sec)
#

Change the block size to the underlying sector size and the command will sail through.
The same also applies to alignment. For HDDs, raw devices need to be read/written on 512-byte boundaries.

20-100 Why are all possible combinations of unit + partition created under /dev even though there's no corresponding device in the machine?

Hysterical raisins, I expect, but, see devpubd.8

kamil

kamil Also, inside the kernel, each block device must be registered on top of an existing character device.

kamil

rvp

see devpubd.8

devpubd(8) is a workaround to the static /dev. Whenever we get some hardware stuff in numbers exceeding the defaults, this daemon, scanning the drvctl database, will create the node for us.

20-100

Thanks a lot for these helpful explanations! 🙂
I'll share my work here when I'll be done.