Messages in this thread | | | From | Thomas Hellstrom <> | Subject | [PATCH RFC 0/2] locking: Fix ww_mutex algorithm inconsistency. | Date | Fri, 25 May 2018 14:38:53 +0200 |
| |
The algorithm used for linux Wound/Wait mutexes, is actually not Wound/Wait but Wait/Die. See for example
http://www.mathcs.emory.edu/~cheung/Courses/554/Syllabus/8-recv+serial/deadlock-compare.html
Rather than renaming them across the tree to something like Wait/Die mutexes or Deadlock Avoidance mutexes, this patch set implements also the Wound/Wait algorithm It shouldn't touch the binary ordinary mutex paths when compiled with otpimization. We use Wound/Wait for the modeset locks which in theory may benefit slightly from Wound/Wait rather than Wait/Die, but Wait/Die is actually superior for larger number of simultaneous contending transactions so we keep the original implementation as a choice, and don't touch the algorithm used by the reservation objects.
Performance- and functional testing has been done using
git://people.freedesktop.org/~thomash/ww_mutex_test
Using WW_BUILTIN and 4 and 16 competing threads.
Another option is of course to ignore the the naming / algorithm inconsistency.
| |