> On Fri, 1 Aug 2025 12:21:05 +0200 > Tomasz Duszynski <tduszyn...@marvell.com> wrote: > > > /proc/sys/kernel/perf_user_access attribute allow user process to access > > perf counters. Though in order to change it binary requires elevated > > capabilities or must be run as root. If that's not the case counter > > access remains disabled. Hence to avoid confusion log message that > > that warns user about that. > > > > Signed-off-by: Tomasz Duszynski <tduszyn...@marvell.com> > > --- > > What do Linux perf tools do in this case?
perf does not collect data directly reading pmu rigisters. It gets data by calling read() on descriptor returned by perf_event_open(). This series on the other hand, is for allowing that direct through-register access to save on cycles. `perf test -v user` may be used to check if direct access can work. On arm64 with perf_user_access=0: $ perf test -v user Couldn't bump rlimit(MEMLOCK), failures may take place when creating BPF maps, etc 4.2: User space counter reading of instructions : Skip (permissions) 4.3: User space counter reading of cycles : Skip (permissions) On arm64 with perf_user_access=1: $ perf test -v user Couldn't bump rlimit(MEMLOCK), failures may take place when creating BPF maps, etc 4.2: User space counter reading of instructions : Ok 4.3: User space counter reading of cycles : Ok On x86 there is no perf_user_access knob available so you achieve same functionality with perf_event_paranoid.
