Safety harnesses (was: Is a successful call to write(2) atomic?)

[Graham Perrin]

On 16/06/2021 14:17, Arthur Chance wrote:

> … More like "we think the builders may have installed the floors in this
> darkened building, but we suggest a flashlight and safety harness" …


Vaguely related, from the perspective of someone who's not a developer, 
I was recently surprised by the potential downsides of defaults when 
FreeBSD is installed to UFS.

For a moment, back to the opening post from Ronald F. Guilmette:

>> … Is a block of data that is successfully written …

The surprise, to me, was losing fifty-something seconds' worth of data 
in a kernel panic situation. Subsequently found, in the FreeBSD Handbook 
under '12.10.2. Soft Updates' 
<https://docs.freebsd.org/en/books/handbook/config/#soft-updates>:

>>> … Soft Updates guarantee file system consistency in the case of a 
>>> crash, but could easily be several seconds or even a minute behind 
>>> updating the physical disk. If the system crashes, unwritten data 
>>> may be lost. …

(I expected _some_ data loss, but sixty seconds surprised me.)

It was suggested that disabling UFS soft updates might improve the 
situation.

Through subsequent tests, with a disposable installation – soft updates 
disabled, carefully timed interruptions whilst using pkg-install(8) – I 
soon produced what seemed to be a wrecked base system. Photographs after 
the crash:

<https://i.imgur.com/YdI3LSJ.png>
<https://i.imgur.com/EpRE8S1.png>

– and after booting from a usable operating system, to check then repair 
the file system:

<https://i.imgur.com/L1YXVgT.png>
<https://i.imgur.com/g1cCy3l.png>
<https://i.imgur.com/ZaM1vvx.png>

With the file system repaired: still, the base system was broken. 
Metaphorically, dropped without a safety harness from a height of more 
than one floor :-)

 From this end result, I assume that:

1. for guaranteed file system consistency with UFS, soft updates may be 
_highly_ desirable, in some situations

2. some other approach should be taken to reducing the potential scope 
(sixty seconds) of data loss.

---

Some surprise at the default delays for syncing files, directories and 
metadata. Respectively: 30, 29 and 28 seconds.

syncer(4) <https://man.freebsd.org/syncer(4)>

----

For the computer where I wrecked the file system, I imagine that (for a 
future test installation) this combination will be reasonable:

- soft updates disabled

- mount(8) option 'sync' set in /etc/fstab (does this reduce the risk of
   wreckage with soft updates disabled?)

- reduced delays for syncer(4).

That's my imagination, although honestly, the whole thing messed with my 
mind.

I'm much happier to simply use OpenZFS (or ZFS).