One needs to Jeff's uncommitted locking changes to see any benefit from this if one touches any pages in the mapping and even then it's only a ~20% win because there is then so much contention on the vm page queue mutex. Jeff is hesitant to sink too much energy in to this on behalf of a microbenchmark. There are clear wins to be had elsewhere without the complexity and risk of VM, so dropping this for now.

LGTM

@lstewart ping

In D15275#322358, @kib wrote:

In D15275#322250, @kmacy wrote:

In D15275#322071, @kib wrote:

How does an event description from the json tables is matched against the index from pmc_events.h ?

It works so long as the FreeBSD version was named correctly. I have an aliases table pmu_utils.c for things like UNHALTED_CORE_CYCLES and LLC_MISSES. If the table lookup fails it will just use the default sampling rate that is used on HEAD. Ultimately, on supported architectures I'd like to switch from using the ad hoc defines in pmc_events.h to using the json tables from Intel, IBM, and Cavium.

So did you verified that the name matches ? What are the plan for non-matching names ?

Hi Andi -
Thanks for the prompt response. I will definitely be taking a look at the upstream version. All I've done so far is add table generation to the build process and use the generated tables to provide sane default sampling rates . https://reviews.freebsd.org/D15275

In D15275#322071, @kib wrote:

How does an event description from the json tables is matched against the index from pmc_events.h ?

@kib Thanks. If I touch a page in the mapping that would be more representative correct? Page faults appear to be globally serialized by the per-domain reservations mutex, so that would likely scale no better than the no-op case without your change or mine.

Actually - @cognet could you please fix man page installation for ck?

It looks like the man pages aren't installed for ck. I'll submit a review for that.

! In D15231#321789, @kib wrote:
Where is the documentation for the ck epoch, both interface and implementation.

@shurd @sbruno might one of you commit this?

@shurd does this still work or should it be abandoned?

We needed to refine this further at ISLN.

@lstewart is this worth updating or should I just abandon?

I don't anticipate finding the time to add USERPTR support any time soon.

remove bogus license tags, the source is clearly BSD from Intel

rebase

Fix compile issues on non tier 1 architectures

Fix various panics under v6 as well as world build issue

• spinlock_{enter, exit} are only gratuitous in functions only called from hardclock.
• add copyright to the two heavily modified files

remove gratuitous spinlock_{enter,exit} calls

• make buffer count a function of the number of cpus
• allocate buffers and per-cpu structures from the right domain
• reduce nsamples to the number required assuming at most ~8 samples per tick

Fix most false sharing in pmc

• check watchdog and reset flags with lock held
• clear watchdog flag

In D15010#316190, @mav wrote:

I am not closely familiar with WITNESS, so just a feeling: the long lists of blessed locks and their combinations promises high chances for them to be forgotten on following ZFS updates.

@markj you can find my attempt at that here https://github.com/mattmacy/networking/tree/projects/witnessfix2 -- you can have a go at it. If you don't have the time / inclination I'd _really_ appreciate it if we didn't let the best become the enemy of the good. If we want stacks of reversals we can comment out blessed entries.

In D15010#316092, @markj wrote:

Why not set WITNESS_BLESSED when the first reversal is detected, and have blessed() return true if WITNESS_BLESSED is set? Then the implementation will still be cheap and we could retain the option of logging the stacks, so investigating the reversals will be easier.

add v6 support and witness update

In D15010#316092, @markj wrote:

Why not set WITNESS_BLESSED when the first reversal is detected, and have blessed() return true if WITNESS_BLESSED is set? Then the implementation will still be cheap and we could retain the option of logging the stacks, so investigating the reversals will be easier.

Don't include extra changes

fix typo

add full context

add mls->mls_lock + dl->dl_lock

In D14967#315442, @erj wrote:

I don't see the console hanging on the version of ixl-iflib where I did see them when watchdogs happened, so this appears to fix the deadlock in that situation.

Split in to many separate reviews.