Dirtied shared pages profiles

[caolanm]

Describe the Bug

Flamegraphs of where shared pages are dirtied

Steps to Reproduce

1. perf probe --add do_wp_page
2. perf record -e probe:do_wp_page -o perf.data.write_shared -F5 --call-graph dwarf,65528 -a sleep 1h
3. perf script -i ./perf.data.write_shared --no-inline | stackcollapse-perf.pl | sed -E -s "s/^kitbroker[^;]+/kitbroker/" | sed -E -s "s/^docbroker[^;]+/docbroker/" | sed -E -s "s/^kit_spare[^;]+/kit_spare/" | flamegraph.pl

Sample initial flamegraph

[mmeeks]

Oooh - interesting =) so it looks like the filter cache is not being too helpful the cppu::iCopyConstructSequence thing - I -guess- is reference fiddling and perturbing the underlying data structure there from FilterCache::getItem I guess.

configmgr::Access::getByHierarchicalName() is also a surprise page dirtier - almost certainly again by bogus referencing of C++ objects: our strings should all be static-ized

Indeed Document::load has a lot of dirtying coming from configmr. @grandinj I thought we had tweaked the objects there to not have mutated references as we read the config ?

[caolanm]

This one is for 12 hours

[grandinj]

Indeed Document::load has a lot of dirtying coming from configmr. @grandinj I thought we had tweaked the objects there to not have mutated references as we read the config ?

There are a couple of places that still use the underlying ref-counting configmgr APIs.

FilterCache is one of them, and the other is the stuff that loads the UI config (toolbars, sidebars, etc)

I don't see any decent path to improving those.