Make DirCache default to 0 on Windows relays, if we can't fix the mmap issues

In #24857 (moved), tor's consensus diff cache implementation causes high CPU load on Windows.

As a workaround, we could make DirCache default to 0 on Windows.

changed milestone to %Tor: 0.4.4.x-final

Trac:
Parent Ticket: #24857 (moved)

added 032-unreached-backport 033-removed-20180326 033-triage-20180326 033-unreached-backport-maybe 034-deferred-20180602 035-removed-20180711 040-roadmap-proposed 044-should component::core tor/tor cpu linux milestone::Tor: 0.4.4.x-final network-health owner::ahf parent::24857 performance points::0.2 postfreeze-ok priority::medium regression severity::normal status::assigned type::defect version::tor 0.3.1.9 windows labels

Having a DirCache 0 can still cause >= 90% CPU for me. But looking at the call-stacks of these CPU-hog threads, I noticed tor_cond_wait() was a candidate.

So just adding a SleepEx() in that function, I no longer see any such high CPU usage (I think after only 5 hours of run-time). Here's my patch:

--- a/lib/thread/compat_winthreads.c 2020-01-24 21:59:07
+++ b/lib/thread/compat_winthreads.c 2020-03-05 14:11:58
@@ -170,6 +170,8 @@
   do {
     DWORD res;
     res = WaitForSingleObject(cond->event, ms);
+
+    SleepEx(1, TRUE);
     EnterCriticalSection(&cond->lock);
     if (cond->n_to_wake &&
         cond->generation != generation_at_start) {

Trac:
Username: gvanem

Hmm. If that helps, then there is probably another bug in the function. Are you saying that this function is busy-looping even when there is nothing to do? If you're looking at this in a debugger, can you tell me the values for ms, cond->n_to_wake, and cond->generation_at_start, when the critical section is entered?

Hm. Now that WinXP is obsolete, maybe we should move this function to use CONDITION_VARIABLE instead...

(I've opened a ticket #33554 (moved) for this CPU issue you're talking about here, since this ticket seems to be about changing the dircache default. If you can answer there, that would be great!)

If this bug-tracker had been on Github, I would have seen your reply sooner. But you insist on using this German Trac crap?

Trac:
Username: gvanem

Replying to gvanem:

If this bug-tracker had been on Github, I would have seen your reply sooner. But you insist on using this German Trac crap?

We're hoping to migrate to a gitlab installation some time this year. Sorry that trac isn't working out for you. Do you think you might be able to answer any of the questions on this or the other ticket?

Add 044-must to all "regression" tickets in 0.4.4

Trac:
Summary: Make DirCache default to 0 on Windows relays to Make DirCache default to 0 on Windows relays, if we can't fix the mmap issues

Mark tickets which are important or safe enough to look at post-freeze for 0.4.4.

Trac:
Keywords: N/A deleted, postfreeze-ok added

Trac:
Owner: N/A to ahf
Status: new to assigned

changed time estimate to 1h 36m

mentioned in issue #33554 (moved)

moved to tpo/core/tor#33411

mentioned in issue tpo/core/tor#33554 (closed)