runtime: express lock rank graph as a DAG like go/build/deps_test.go

[aclements]

The runtime's lock rank graph currently requires a full manual transitive closure of the lock rank graph. This makes it annoying to maintain and some of the edge lists are very long.

We should replace this with something like the DAG language in go/build/deps_test.go. In particular, the runtime's lock ranking falls into a few natural strata, which are very natural to express in the deps DAG, and extremely verbose in the current approach. This would also significantly improve the readability of diffs to the graph (see, for example, 482669d). Finally, it would let us auto-generate the total order, rather than trying to keep them consistent by hand.

We could parse this at runtime startup (only if lock rank checking is enabled anyway), though the current totally static structure has the advantage of being available for lock checking from the moment the process starts.

It may be better to go:generate the static ranking structure from the DAG language. This would keep the structure available immediately at startup, and would also make it easy to share the parser with deps_test.go without trying to fit it into the constraints of early runtime initialization.

[aclements]

I need to clean this up by hand, but here's the current lock graph automatically converted to deps language and reduced:

NONE < dummy, sysmon, sweepWaiters, assistQueue, cpuprof, sweep, pollDesc, deadlock, itab, notifyList, root, rwmutexW, defer, newmHandoff, debugPtrmask, faketimeState, ticks, raceFini, pollCache, debug;
sysmon < scavenge, forcegc;
assistQueue, cpuprof, forcegc, pollDesc, scavenge, sweep, sweepWaiters < sched;
sched < allg, allp, gFree;
allp < timers;
itab < reflectOffs;
scavenge, sweep < hchan;
scavenge < traceBuf;
allg, hchan, reflectOffs, timers, traceBuf < fin;
traceBuf < traceStrings;
allg, hchan, notifyList, reflectOffs, timers, traceStrings < mspanSpecial, profInsert, profBlock, profMemActive, gcBitsArenas, spanSetSpine, mheapSpecial;
profMemActive < profMemFuture;
hchan, root, sched, traceStrings < trace;
fin, notifyList, trace < traceStackTab;
timers < netpollInit;
rwmutexW, sysmon < rwmutexR;
gcBitsArenas, netpollInit, profBlock, profInsert, profMemFuture, spanSetSpine, traceStackTab < gscan;
gscan, rwmutexR < stackpool;
gscan < stackLarge, hchanLeaf;
hchan, notifyList < sudog;
defer, gscan, mspanSpecial, sudog < wbufSpans;
stackLarge, stackpool, wbufSpans < mheap;
mheap, mheapSpecial < globalAlloc;
mheap < pageAllocScav;
deadlock < panic;

To the surprise of no-one, the current lock rank graph is not transitively closed, meaning there are locks for which it's okay to acquire a then b or b then c, but not a then c. The new deps-based approach automatically constructs the full reachability matrix, which will fix this, too.

[gopherbot]

Change https://go.dev/cl/418716 mentions this issue: runtime: add missing trace lock edges

[gopherbot]

Change https://go.dev/cl/418717 mentions this issue: runtime: add mayAcquire annotation for finlock

[gopherbot]

Change https://go.dev/cl/418714 mentions this issue: runtime: delete unused lock ranks

[gopherbot]

Change https://go.dev/cl/418715 mentions this issue: runtime: generate the lock ranking from a DAG description

[gopherbot]

Change https://go.dev/cl/418593 mentions this issue: internal/dag: add a Graph type and make node order deterministic

[gopherbot]

Change https://go.dev/cl/418718 mentions this issue: runtime: make the lock rank DAG make more sense

[gopherbot]

Change https://go.dev/cl/418719 mentions this issue: runtime: clean up panic and deadlock lock ranks