Optimize the vm object bypass and collapse counters

[alc]

Summary

Make the vm object bypass and collapse counters per-CPU counters.

[alc]

vm/vm_object.c

151 ↗

(On Diff #36980)

I probably don't need EARLY_COUNTER.

166 ↗

(On Diff #36980)

I based the ordering, SI_SUB_CPU and SI_ORDER_ANY, on that found in vm_meter.c. I have no idea if this ordering is optimal.

[kib]

vm/vm_object.c

166 ↗

(On Diff #36980)

I think it does not matter, collapses and bypasses start to occur only when forks are executed, which does not happen earlier than init startup. You used the same values as vmcounter_startup(), which is fine.

[alc]

vm/vm_object.c

153 ↗

(On Diff #36980)

Should I put the description on its own line?

[kib]

vm/vm_object.c

153 ↗

(On Diff #36980)

Description on the new line is bde' preferred style. IMO it unnecessarily spends vertical space if the description fits into line length.

[alc]

As an aside, it seems to me that sys/counter.h and/or the man page ought to provide guidance on when counter_u64_alloc() becomes possible. It appears to me that it is earlier than SI_SUB_CPU, specifically, we need the following to have completed:

SYSINIT(pcpu_zones, SI_SUB_KMEM, SI_ORDER_ANY, pcpu_zones_startup, NULL);

[markj]

In D13611#284781, @alc wrote:

As an aside, it seems to me that sys/counter.h and/or the man page ought to provide guidance on when counter_u64_alloc() becomes possible. It appears to me that it is earlier than SI_SUB_CPU, specifically, we need the following to have completed:

SYSINIT(pcpu_zones, SI_SUB_KMEM, SI_ORDER_ANY, pcpu_zones_startup, NULL);

I believe @glebius has some plans to allow earlier creation of per-CPU counters.