commit 18d2048fda76cc1b3715fd2f1a0f6ac3e381d26d
author Adrian Danis <adanis@google.com> Fri Jun 07 07:01:30 2019 +0000
committer CQ bot account: commit-bot@chromium.org <commit-bot@chromium.org> Fri Jun 07 07:01:30 2019 +0000
tree 96dd24164544017e103fecf41a066341ae4a5918
parent ff23a98aa57283652e924f1a8c62e60c0218df8a

[kernel] Count number of spinlocks held

Adds a percpu counter for how many spinlocks are held. This counter is
asserted against during a reschedule to ensure that no spinlocks, aside
from the thread_lock, are held across a thread switch.

For x86 the percpu struct is retrieved and then used for retrieving
both the cpu_num and manipulating num_spinlocks. This is faster than
performing all operations indirectly through gs. For arm there was no
benefit measured from locally caching the percpu pointer and so the
generic READ+WRITE wrappers are used.

The in kernel micro benchmarks (k bench) shows the cycles to perform a
spinlock acquire+release changes slightly:
NUC 47 -> 48
VIM2 41 -> 46

To measure whether there is any noticable performance degredation the
two zircon micro benchmark suites (zircon/perftest,
garnet/zircon_benchmarks) were run on the vim2 and nuc. All tests were
run 5 times with reboots in between and compared with the perfcompare
tool, which found no significant differences.

ZX-147 #comment

Change-Id: I40f15bf3dbe4e4e4219e8a461507369a0092fd7c