memlat: Read perf counters in parallel and reduce system jitterHEADs12.1

Sending synchronous IPIs to other CPUs involves spinning with preemption
disabled in order to wait for each IPI to finish. Keeping preemption off
for long periods of time like this is bad for system jitter, not to mention
the perf event IPIs are sent and flushed one at a time for each event for
each CPU rather than all at once for all the CPUs.

Since the way perf events are currently read is quite naive, rewrite it to
make it exploit parallelism and go much faster. IPIs for reading each perf
event are now sent to all CPUs asynchronously so that each CPU can work on
reading the events in parallel, and the dispatching CPU now sleeps rather
than spins when waiting for the IPIs to finish. Before the dispatching CPU
starts waiting though, it works on reading events for itself and then
reading events which can be read from any CPU in order to derive further
parallelism, and then waits for the IPIs to finish afterwards if they
haven't already.

Furthermore, there's now only one IPI sent to read all of a CPU's events
rather than an IPI sent for reading each event, which significantly speeds
up the event reads and reduces the number of IPIs sent.

This also checks for active SCM calls on a per-CPU basis rather than a
global basis so that unrelated CPUs don't get their counter reads skipped
and so that some CPUs can still receive fresh counter readings.

Overall, this makes the memlat driver much faster and more efficient, and
eliminates significant system jitter previously caused by IPI abuse.

Signed-off-by: Sultan Alsawaf <sultan@kerneltoast.com>
Signed-off-by: Adithya R <gh0strider.2k18.reborn@gmail.com>

0 files changed, 0 insertions, 0 deletions