59 bytes, text/x-review-board-request
ReentrantMonitors are major overkill for what string bundles are doing. They're never used as monitors, and the sections where they are held are guaranteed to never be re-entrant. We should be using simple mutexes instead, since they use significantly less memory and are significantly more performant.
Comment on attachment 8993864 [details] Bug 1473135: Stop using ReentrantMonitor for string bundle mutexes. https://reviewboard.mozilla.org/r/258536/#review265576 Hmm, the patch looks good, but I'm surprised to hear that Monitor is noticably heavier than Mutex. Is there any documentation on mozilla's monitor/mutex implementations and when which one should be used?
Attachment #8993864 - Flags: review?(gandalf) → review+
(In reply to Zibi Braniecki [:gandalf][:zibi] from comment #2) > Hmm, the patch looks good, but I'm surprised to hear that Monitor is > noticably heavier than Mutex. As a rule, every monitor uses at least one mutex, in addition to a cond, which is at least as heavy as a mutex plus the wait list it has to maintain. Re-entrant monitors are generally even worse, but ours are especially bad. They just wrap PRMonitor, which contains a mutex, two conds, an owner thread, a name pointer, and a bunch of accounting data: https://searchfox.org/mozilla-central/rev/ad36eff63e208b37bc9441b91b7cea7291d82890/nsprpub/pr/include/private/primpl.h#1448-1483 They're something like 224 bytes a piece, not including any private data that libc allocates for them. > Is there any documentation on mozilla's monitor/mutex implementations and > when which one should be used? Unfortunately, no. The general rule is to just always use the lightest one you can get away with.
https://hg.mozilla.org/integration/mozilla-inbound/rev/9ba70b8ff0f755d5575333a71267805e0ece02f6 Bug 1473135: Stop using ReentrantMonitor for string bundle mutexes. r=gandalf
You need to log in before you can comment on or make changes to this bug.