neirac - thanks, I'll take a look today.

one other thing i've been thinking about, but haven't had a chance to actually mess with yet, is adding more memory awareness (for lack of a better term) into the zio pipeline

e.g. zfs will happily keep allocating (potentially large) zios and stuff them into the pipeline until it exhausts memory

but a decent amount of them are flagged 'can fail', so one thought I had was when hit some low water mark in free memory (e.g. freemem < lotsfree)

have it just fail zios that are marked 'can fail' instead of sending them out to disk

and maybe something to throttle things on the 'frontend' of the zio pipeline when we're low on memory (which might make it easier to allow swap related activity to pass through since at that stage, it's a lot easier to flag/identify 'this is swapping related')

though I don't know that it'd be a true 'fix' as much as (hopefully) make it a lot harder to get into the death spiral

it can be particularly bad if your have very slow drives (e.g. VM with very busy backends can really throw zfs into fits)

[illumos-gate] 15980 sort should not be isaexec'd -- Richard Lowe <richlowe⊙rn>