honest manual: BUGS This program does not work very well.

anyhow, I did a small test, picked MOS DVA, nuked with /dev/zero block pointed by dva[0], observed loader getting checksum error and picking dva[1] instead, but also zpool import poolname was all happy - the only bug was that self repair event is not recorded for it -- the block was repaired silently, but zpool status does not reflect the checksum error

tsoome: could we make loader say something if it has to read non-primary blocks?

or is that a thing were just supposed to silently succeed? I reckon I'd probably want to know

sure we can, currently is just walking through dva list and is only complaining when all reads have failed.

and yep, its really about noise...

tsoome: yeah, I'm never good at balancing noise v. action

but I think anything weird about my storage I'd want a clue as early as possible, just not it spammed

(for eg, my windows experience when disks fail is most of the I/O that fails is the windows event logs saying the disk is failing...)

I actually do think it is good idea.

richlowe: periodic scrubs should catch and repair that sort of thing.

(loader presumably doesn't attempt to repair the busted copy of the block..)

no, it does not.

maybe the way to go is for the loader to remember that this happened and somehow leave a note for the kernel. "drop this into the FMA error log and let the ZFS diagnosis engine do something about it.."

(boot module of nvlist?)

i do have a change to improve sd + zfs.. at least for supported SCSI disks (enterprise ones seem to) where a read fails due to a bad block will trigger remapping of the bad block and then a zfs self-heal

(today, zfs self healing only works if the drive returns a _successful_ I/O with invalid data)

but i've been working on a userland iscsi client (based off a rust crate) to simulate things for testing

sommerfeld nice one;)

jbk: yeah, having simulated hardware that can fail on demand is crucial for things like that..

just E810 is like a bottomless pit of code

i thought i was almost done, and then 'no, to get RSS working so more than one RX ring is actually used, it requires thousands of lines of additional code'

(20k lines to date)

ow.

[illumos-gate] Backout: 18149 ld(1) should support -zsysroot -- Richard Lowe <richlowe⊙rn>

I think a bare minimum of you were in a position to experience that, but if you were you have my apologies

been in tree since August 27th?

I'm told that in practice it doesn't happen through the gcc frontends, you would have to have that commit and be using (or possibly building?) clang