* [RFC] arm64: perf: associate LL with L2 cache accesses and refills @ 2015-11-04 11:24 hw.claudio at gmail.com 2015-11-04 11:39 ` Mark Rutland 0 siblings, 1 reply; 3+ messages in thread From: hw.claudio at gmail.com @ 2015-11-04 11:24 UTC (permalink / raw) To: linux-arm-kernel From: Claudio Fontana <claudio.fontana@huawei.com> Signed-off-by: Claudio Fontana <claudio.fontana@huawei.com> Cc: Ammar Saeed <ammar.saeed@huawei.com> --- Hello, as part of some experiments with the Juno ARM64 board, we needed to get readings from the PMU regarding L2 Cache hits and misses, but we noticed that the L2 Cache Access and Refill performance counters were not hooked up in the perf API. We just did that, and that seems to produce correct results on the Juno. However I guess that these registers are not hooked up by default due to differences between different boards...how could this be done taking account of the different possible implementations? I send this as an initial RFC to try to kickoff discussion about this. Thank you, Claudio Fontana arch/arm64/kernel/perf_event.c | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/arch/arm64/kernel/perf_event.c b/arch/arm64/kernel/perf_event.c index f9a74d4..f72f2ff 100644 --- a/arch/arm64/kernel/perf_event.c +++ b/arch/arm64/kernel/perf_event.c @@ -728,6 +728,11 @@ static const unsigned armv8_pmuv3_perf_cache_map[PERF_COUNT_HW_CACHE_MAX] [C(L1D)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_L1_DCACHE_ACCESS, [C(L1D)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV8_PMUV3_PERFCTR_L1_DCACHE_REFILL, + [C(LL)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_L2_CACHE_ACCESS, + [C(LL)][C(OP_READ)][C(RESULT_MISS)] = ARMV8_PMUV3_PERFCTR_L2_CACHE_REFILL, + [C(LL)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_L2_CACHE_ACCESS, + [C(LL)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV8_PMUV3_PERFCTR_L2_CACHE_REFILL, + [C(BPU)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_PC_BRANCH_PRED, [C(BPU)][C(OP_READ)][C(RESULT_MISS)] = ARMV8_PMUV3_PERFCTR_PC_BRANCH_MIS_PRED, [C(BPU)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_PC_BRANCH_PRED, -- 1.8.5.3 ^ permalink raw reply related [flat|nested] 3+ messages in thread
* [RFC] arm64: perf: associate LL with L2 cache accesses and refills 2015-11-04 11:24 [RFC] arm64: perf: associate LL with L2 cache accesses and refills hw.claudio at gmail.com @ 2015-11-04 11:39 ` Mark Rutland 2015-11-04 12:50 ` Claudio Fontana 0 siblings, 1 reply; 3+ messages in thread From: Mark Rutland @ 2015-11-04 11:39 UTC (permalink / raw) To: linux-arm-kernel On Wed, Nov 04, 2015 at 12:24:13PM +0100, hw.claudio at gmail.com wrote: > From: Claudio Fontana <claudio.fontana@huawei.com> > > Signed-off-by: Claudio Fontana <claudio.fontana@huawei.com> > Cc: Ammar Saeed <ammar.saeed@huawei.com> > --- > > Hello, Hi, > as part of some experiments with the Juno ARM64 board, we needed to get > readings from the PMU regarding L2 Cache hits and misses, but we noticed > that the L2 Cache Access and Refill performance counters were not hooked > up in the perf API. We just did that, and that seems to produce correct > results on the Juno. > > However I guess that these registers are not hooked up by default due to > differences between different boards...how could this be done taking > account of the different possible implementations? The events we list for PMUv3 are those which are required to be implemented (see "D5.10.6 Required events" in ARM DDI 0487A.h). All others (including the L2 events you add) are optional and may or may not be implemented, so we can't expose them for all PMUv3 implementations. To account for different events, we will shortly be exposing separate logical PMUs (see [1]), which will allow us to support each CPU's set of supported events independently. That's queued in the arm64 tree [2] currently. I see that per their respective TRMs, both Cortex-A53 and Cortex-A57 support these L2 events. It looks like when I added specialised support [3,4] I simply missed them. Fancy sending a patch to correct that? Thanks, Mark. [1] http://lists.infradead.org/pipermail/linux-arm-kernel/2015-October/374053.html [2] https://git.kernel.org/cgit/linux/kernel/git/arm64/linux.git/log/?h=for-next/core [3] http://lists.infradead.org/pipermail/linux-arm-kernel/2015-October/374052.html [4] http://lists.infradead.org/pipermail/linux-arm-kernel/2015-October/374056.html > I send this as an initial RFC to try to kickoff discussion about this. > > Thank you, > > Claudio Fontana > > arch/arm64/kernel/perf_event.c | 5 +++++ > 1 file changed, 5 insertions(+) > > diff --git a/arch/arm64/kernel/perf_event.c b/arch/arm64/kernel/perf_event.c > index f9a74d4..f72f2ff 100644 > --- a/arch/arm64/kernel/perf_event.c > +++ b/arch/arm64/kernel/perf_event.c > @@ -728,6 +728,11 @@ static const unsigned armv8_pmuv3_perf_cache_map[PERF_COUNT_HW_CACHE_MAX] > [C(L1D)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_L1_DCACHE_ACCESS, > [C(L1D)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV8_PMUV3_PERFCTR_L1_DCACHE_REFILL, > > + [C(LL)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_L2_CACHE_ACCESS, > + [C(LL)][C(OP_READ)][C(RESULT_MISS)] = ARMV8_PMUV3_PERFCTR_L2_CACHE_REFILL, > + [C(LL)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_L2_CACHE_ACCESS, > + [C(LL)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV8_PMUV3_PERFCTR_L2_CACHE_REFILL, > + > [C(BPU)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_PC_BRANCH_PRED, > [C(BPU)][C(OP_READ)][C(RESULT_MISS)] = ARMV8_PMUV3_PERFCTR_PC_BRANCH_MIS_PRED, > [C(BPU)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_PC_BRANCH_PRED, > -- > 1.8.5.3 > ^ permalink raw reply [flat|nested] 3+ messages in thread
* [RFC] arm64: perf: associate LL with L2 cache accesses and refills 2015-11-04 11:39 ` Mark Rutland @ 2015-11-04 12:50 ` Claudio Fontana 0 siblings, 0 replies; 3+ messages in thread From: Claudio Fontana @ 2015-11-04 12:50 UTC (permalink / raw) To: linux-arm-kernel On 04.11.2015 12:39, Mark Rutland wrote: > On Wed, Nov 04, 2015 at 12:24:13PM +0100, hw.claudio at gmail.com wrote: >> From: Claudio Fontana <claudio.fontana@huawei.com> >> >> Signed-off-by: Claudio Fontana <claudio.fontana@huawei.com> >> Cc: Ammar Saeed <ammar.saeed@huawei.com> >> --- >> >> Hello, > > Hi, > >> as part of some experiments with the Juno ARM64 board, we needed to get >> readings from the PMU regarding L2 Cache hits and misses, but we noticed >> that the L2 Cache Access and Refill performance counters were not hooked >> up in the perf API. We just did that, and that seems to produce correct >> results on the Juno. >> >> However I guess that these registers are not hooked up by default due to >> differences between different boards...how could this be done taking >> account of the different possible implementations? > > The events we list for PMUv3 are those which are required to be > implemented (see "D5.10.6 Required events" in ARM DDI 0487A.h). All > others (including the L2 events you add) are optional and may or may not > be implemented, so we can't expose them for all PMUv3 implementations. > > To account for different events, we will shortly be exposing separate > logical PMUs (see [1]), which will allow us to support each CPU's set > of supported events independently. That's queued in the arm64 tree [2] > currently. > > I see that per their respective TRMs, both Cortex-A53 and Cortex-A57 > support these L2 events. It looks like when I added specialised support > [3,4] I simply missed them. Fancy sending a patch to correct that? > > Thanks, > Mark. I gave a first look at the resources you provided, I am looking at the for-next/core branch you mentioned. However, when reading the Cortex-A-53 manual it seems that even for those specific CPUs the L2 Counters are optional, as the L2 Cache itself is optional. Quoting from "12.4.2 Performance Monitors Common Event Identification Register 0": Table 12-6 on page 12-10 shows the PMCEID0_EL0 bit assignments [23] 0x17 L2D_CACHE_REFILL L2 Data cache refill: 0 This event is not implemented if the Cortex-A53 processor has been configured without an L2 cache. 1 This event is implemented if the Cortex-A53 processor has been configured with an L2 cache. [22] 0x16 L2D_CACHE L2 Data cache access: 0 This event is not implemented if the Cortex-A53 processor has been configured without an L2 cache. 1 This event is implemented if the Cortex-A53 processor has been configured with an L2 cache. I don't see that we are reading this register to check whether the hardware supports those counters or not.. shouldn't we? However, I think it should be a direct consequence of L2 cache being present, so maybe we can use the existing struct cpu_cacheinfo "num_levels"? For A-57 it does not seem to be an issue since as far as I can see from the manual, the L2 cache is always present. Do I understand this correctly? Ciao, Claudio > > [1] http://lists.infradead.org/pipermail/linux-arm-kernel/2015-October/374053.html > [2] https://git.kernel.org/cgit/linux/kernel/git/arm64/linux.git/log/?h=for-next/core > [3] http://lists.infradead.org/pipermail/linux-arm-kernel/2015-October/374052.html > [4] http://lists.infradead.org/pipermail/linux-arm-kernel/2015-October/374056.html > >> I send this as an initial RFC to try to kickoff discussion about this. >> >> Thank you, >> >> Claudio Fontana >> >> arch/arm64/kernel/perf_event.c | 5 +++++ >> 1 file changed, 5 insertions(+) >> >> diff --git a/arch/arm64/kernel/perf_event.c b/arch/arm64/kernel/perf_event.c >> index f9a74d4..f72f2ff 100644 >> --- a/arch/arm64/kernel/perf_event.c >> +++ b/arch/arm64/kernel/perf_event.c >> @@ -728,6 +728,11 @@ static const unsigned armv8_pmuv3_perf_cache_map[PERF_COUNT_HW_CACHE_MAX] >> [C(L1D)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_L1_DCACHE_ACCESS, >> [C(L1D)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV8_PMUV3_PERFCTR_L1_DCACHE_REFILL, >> >> + [C(LL)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_L2_CACHE_ACCESS, >> + [C(LL)][C(OP_READ)][C(RESULT_MISS)] = ARMV8_PMUV3_PERFCTR_L2_CACHE_REFILL, >> + [C(LL)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_L2_CACHE_ACCESS, >> + [C(LL)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV8_PMUV3_PERFCTR_L2_CACHE_REFILL, >> + >> [C(BPU)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_PC_BRANCH_PRED, >> [C(BPU)][C(OP_READ)][C(RESULT_MISS)] = ARMV8_PMUV3_PERFCTR_PC_BRANCH_MIS_PRED, >> [C(BPU)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_PC_BRANCH_PRED, >> -- >> 1.8.5.3 >> ^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2015-11-04 12:50 UTC | newest] Thread overview: 3+ messages (download: mbox.gz follow: Atom feed -- links below jump to the message on this page -- 2015-11-04 11:24 [RFC] arm64: perf: associate LL with L2 cache accesses and refills hw.claudio at gmail.com 2015-11-04 11:39 ` Mark Rutland 2015-11-04 12:50 ` Claudio Fontana
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox; as well as URLs for NNTP newsgroup(s).