On 4/14/2022 12:09 PM, Ian Rogers wrote: > On Wed, Apr 13, 2022 at 11:17 AM Liang, Kan wrote: >> >> >> >> On 4/13/2022 1:09 PM, Ian Rogers wrote: >>> On Wed, Apr 13, 2022 at 9:37 AM Liang, Kan wrote: >>>> >>>> >>>> >>>> On 4/13/2022 12:03 PM, Ian Rogers wrote: >>>>> 3) Weak group doesn't fall back to no group: >>>> >>>> That's because the group validation code doesn't take pinned events, >>>> such as the NMI watchdog, into account. >>>> >>>> I proposed a kernel patch to fix it, but it's rejected. It should be >>>> hard to find a generic way to fix it from the kernel side. >>>> https://lore.kernel.org/lkml/1565977750-76693-1-git-send-email-kan.liang@linux.intel.com/ >>>> >>>> Maybe we can workaround it from the perf tool side? >>>> For example, for each weak group with cycles event and NMI watchdog is >>>> enabled, add an extra cycles event when opening the group. If the open >>>> fails with the extra cycles event, fall back to no group. After the >>>> extra cycles event check, remove the extra cycles. >>>> >>>> What do you think? >>> >>> Thanks Kan, it is a shame the kernel support is lacking here. I'm not >>> sure what you are proposing for the perf tool to do. So: >>> >>>> for each weak group with cycles event and NMI watchdog >>> >>> Okay, let's try Branching_Overhead as mentioned in this report - but >>> the event is CPU_CLK_UNHALTED.THREAD here :-/ >>> >>>> add an extra cycles event when opening the group >>> >>> So the perf_event_open doesn't fail here for me: >>> $ perf stat -e '{BR_INST_RETIRED.NEAR_CALL,BR_INST_RETIRED.NEAR_TAKEN,BR_INST_RETIRED.NOT_TAKEN,BR_INST_RETIRED.CONDITIONAL,CPU_CLK_UNHALTED.THREAD},cycles' >>> -a sleep 1 >>> >> >> No, I mean modifying the perf tool code and add an extra cycles in the >> weak group. >> >> Here is a very initial POC patch, which should prove the idea. > > So I was unaware of this behavior, great find! However, it seems > difficult to exploit. Here is the extra cycles "fixing" a weak group: > ``` > $ perf stat -e '{BR_INST_RETIRED.NEAR_CALL,BR_INST_RETIRED.NEAR_TAKEN,BR_INST_RETIRED.NOT_TAKEN,BR_INST_RETIRED.CONDITIONAL,cycles,cycles}:W' > -a sleep 1 > > Performance counter stats for 'system wide': > > 18,782,301 BR_INST_RETIRED.NEAR_CALL > (66.64%) > 153,325,072 BR_INST_RETIRED.NEAR_TAKEN > (66.64%) > 75,443,263 BR_INST_RETIRED.NOT_TAKEN > (66.64%) > 156,568,769 BR_INST_RETIRED.CONDITIONAL > (66.66%) > 1,870,812,571 cycles > (66.72%) > 1,890,508,326 cycles > (66.70%) > > 1.006371081 seconds time elapsed > ``` > > But if the original group has 1 less counter we will fail with the > duplicate cycles: I should use close() instead of evsel__close() in the POC, since I didn't use the evsel__open() to open the extra cycles event. evsel->core.fd is always NULL. evsel__close() doesn't close the extra cycles event. Please apply the below change to the POC patch. if (fd < 0) { evlist__reset_weak_group(evsel_list, counter, false); second_pass = true; } else { + close(fd); - evsel__close(evsel); } [kan@kanliang-dev perf]$ sudo ./perf stat -e '{BR_INST_RETIRED.NEAR_CALL,BR_INST_RETIRED.NEAR_TAKEN,BR_INST_RETIRED.NOT_TAKEN,CPU_CLK_UNHALTED.THREAD}:W' -C0 sleep 1 Performance counter stats for 'CPU(s) 0': 570,963 BR_INST_RETIRED.NEAR_CALL 2,271,463 BR_INST_RETIRED.NEAR_TAKEN 1,942,059 BR_INST_RETIRED.NOT_TAKEN 19,938,075 CPU_CLK_UNHALTED.THREAD 1.002352851 seconds time elapsed [kan@kanliang-dev perf]$ sudo ./perf stat -e '{BR_INST_RETIRED.NEAR_CALL,BR_INST_RETIRED.NEAR_TAKEN,BR_INST_RETIRED.NOT_TAKEN,BR_INST_RETIRED.CONDITIONAL,CPU_CLK_UNHALTED.THREAD}:W' -C0 sleep 1 Performance counter stats for 'CPU(s) 0': 563,096 BR_INST_RETIRED.NEAR_CALL (79.95%) 2,077,100 BR_INST_RETIRED.NEAR_TAKEN (79.97%) 1,602,897 BR_INST_RETIRED.NOT_TAKEN (80.05%) 2,332,397 BR_INST_RETIRED.CONDITIONAL (80.05%) 19,255,825 CPU_CLK_UNHALTED.THREAD (79.99%) 1.002250279 seconds time elapsed [kan@kanliang-dev perf]$ sudo ./perf stat -e '{BR_INST_RETIRED.NEAR_CALL,BR_INST_RETIRED.NEAR_TAKEN,BR_INST_RETIRED.NOT_TAKEN,BR_INST_RETIRED.CONDITIONAL,CPU_CLK_UNHALTED.THREAD,cycles}:W' -C0 sleep 1 Performance counter stats for 'CPU(s) 0': 633,648 BR_INST_RETIRED.NEAR_CALL (66.67%) 2,376,309 BR_INST_RETIRED.NEAR_TAKEN (66.68%) 1,380,715 BR_INST_RETIRED.NOT_TAKEN (66.68%) 2,064,048 BR_INST_RETIRED.CONDITIONAL (66.68%) 19,493,096 CPU_CLK_UNHALTED.THREAD (66.66%) 19,396,279 cycles (66.64%) 1.002327588 seconds time elapsed > ``` > $ perf stat -e '{BR_INST_RETIRED.NEAR_CALL,BR_INST_RETIRED.NEAR_TAKEN,BR_INST_RETIRED.NOT_TAKEN,cycles,cycles}:W' > -a sleep 1 > Performance counter stats for 'system wide': > > BR_INST_RETIRED.NEAR_CALL > (0.00%) > BR_INST_RETIRED.NEAR_TAKEN > (0.00%) > BR_INST_RETIRED.NOT_TAKEN > (0.00%) > cycles > (0.00%) > cycles > (0.00%) > > 1.005599088 seconds time elapsed > > Some events weren't counted. Try disabling the NMI watchdog: > echo 0 > /proc/sys/kernel/nmi_watchdog > perf stat ... > echo 1 > /proc/sys/kernel/nmi_watchdog > The events in group usually have to be from the same PMU. Try > reorganizing the group. > ``` > > If we add two extra cycles or the original group is smaller then it is "fixed": > ``` > $ perf stat -e '{BR_INST_RETIRED.NEAR_CALL,BR_INST_RETIRED.NEAR_TAKEN,BR_INST_RETIRED.NOT_TAKEN,cycles}:W' > -a sleep 1 > > Performance counter stats for 'system wide': > > 20,378,789 BR_INST_RETIRED.NEAR_CALL > 168,420,963 BR_INST_RETIRED.NEAR_TAKEN > 96,330,608 BR_INST_RETIRED.NOT_TAKEN > 1,652,230,042 cycles > > 1.008757590 seconds time elapsed > > $ perf stat -e '{BR_INST_RETIRED.NEAR_CALL,BR_INST_RETIRED.NEAR_TAKEN,BR_INST_RETIRED.NOT_TAKEN,cycles,cycles,cycles}:W' > -a sleep 1 > > Performance counter stats for 'system wide': > > 37,696,638 BR_INST_RETIRED.NEAR_CALL > (66.62%) > 298,535,151 BR_INST_RETIRED.NEAR_TAKEN > (66.63%) > 297,011,663 BR_INST_RETIRED.NOT_TAKEN > (66.63%) > 3,155,711,474 cycles > (66.65%) > 3,194,919,959 cycles > (66.74%) > 3,126,664,102 cycles > (66.72%) > > 1.006237962 seconds time elapsed > ``` > > So the extra cycles is needed to fix weak groups when the nmi watchdog > is enabled and the group is an architecture dependent size. Yes, the size of the group depends on the architecture, but perf tool doesn't need to know the HW details. For this case, perf tool just sends the request with an extra cycles event in the group and lets kernel decide. Thanks, Kan > > Thanks, > Ian > >> diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c >> index b7fe88beb584..782c3d7f1b32 100644 >> --- a/tools/perf/builtin-stat.c >> +++ b/tools/perf/builtin-stat.c >> @@ -71,7 +71,9 @@ >> #include "util/bpf_counter.h" >> #include "util/iostat.h" >> #include "util/pmu-hybrid.h" >> +#include "util/util.h" >> #include "asm/bug.h" >> +#include "perf-sys.h" >> >> #include >> #include >> @@ -777,6 +779,8 @@ static enum counter_recovery >> stat_handle_error(struct evsel *counter) >> return COUNTER_FATAL; >> } >> >> +#define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y)) >> + >> static int __run_perf_stat(int argc, const char **argv, int run_idx) >> { >> int interval = stat_config.interval; >> @@ -793,6 +797,7 @@ static int __run_perf_stat(int argc, const char >> **argv, int run_idx) >> struct affinity saved_affinity, *affinity = NULL; >> int err; >> bool second_pass = false; >> + bool has_cycles = false; >> >> if (forks) { >> if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, >> workload_exec_failed_signal) < 0) { >> @@ -821,6 +826,8 @@ static int __run_perf_stat(int argc, const char >> **argv, int run_idx) >> evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) { >> counter = evlist_cpu_itr.evsel; >> >> + if (counter->core.attr.config == 0x3c) >> + has_cycles = true; >> /* >> * bperf calls evsel__open_per_cpu() in bperf__load(), so >> * no need to call it again here. >> @@ -867,6 +874,24 @@ static int __run_perf_stat(int argc, const char >> **argv, int run_idx) >> counter->supported = true; >> } >> >> + //make it model specific. need to move to a better place >> + if (counter->supported && !second_pass && has_cycles && >> counter->weak_group && sysctl__nmi_watchdog_enabled()) { >> + struct evsel *leader = evsel__leader(counter); >> + int group_fd = FD(leader, 0, 0); >> + struct evsel *evsel; >> + int fd; >> + >> + evsel = evsel__new_cycles(0, PERF_TYPE_HARDWARE, >> PERF_COUNT_HW_CPU_CYCLES); >> + fd = sys_perf_event_open(&evsel->core.attr, -1, 0, group_fd, 0x8); >> + >> + if (fd < 0) { >> + evlist__reset_weak_group(evsel_list, counter, false); >> + second_pass = true; >> + } else { >> + evsel__close(evsel); >> + } >> + } >> + >> if (second_pass) { >> /* >> * Now redo all the weak group after closing them, >> >> With the above patch, >> >> $ sudo ./perf stat -e >> '{BR_INST_RETIRED.NEAR_CALL,BR_INST_RETIRED.NEAR_TAKEN,BR_INST_RETIRED.NOT_TAKEN,BR_INST_RETIRED.CONDITIONAL,CPU_CLK_UNHALTED.THREAD}:W' >> -C0 sleep 1 >> >> Performance counter stats for 'CPU(s) 0': >> >> 913,369 BR_INST_RETIRED.NEAR_CALL >> (79.95%) >> 3,648,433 BR_INST_RETIRED.NEAR_TAKEN >> (80.00%) >> 2,481,976 BR_INST_RETIRED.NOT_TAKEN >> (80.05%) >> 3,696,298 BR_INST_RETIRED.CONDITIONAL >> (80.04%) >> 27,792,053 CPU_CLK_UNHALTED.THREAD >> (79.96%) >> >> 1.002224709 seconds time elapsed >> >> >> Thanks, >> Kan