Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752572AbdHQHpK (ORCPT ); Thu, 17 Aug 2017 03:45:10 -0400 Received: from mail-it0-f46.google.com ([209.85.214.46]:36352 "EHLO mail-it0-f46.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752524AbdHQHpG (ORCPT ); Thu, 17 Aug 2017 03:45:06 -0400 MIME-Version: 1.0 In-Reply-To: <1502311335-61580-1-git-send-email-kan.liang@intel.com> References: <1502311335-61580-1-git-send-email-kan.liang@intel.com> From: Stephane Eranian Date: Thu, 17 Aug 2017 00:45:04 -0700 Message-ID: Subject: Re: [PATCH V3] perf: Add PERF_SAMPLE_PHYS_ADDR To: "Liang, Kan" Cc: Peter Zijlstra , "mingo@redhat.com" , LKML , Arnaldo Carvalho de Melo , Jiri Olsa , Thomas Gleixner , "ak@linux.intel.com" Content-Type: text/plain; charset="UTF-8" Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 8105 Lines: 198 On Wed, Aug 9, 2017 at 1:42 PM, wrote: > From: Kan Liang > > For understanding how the workload maps to memory channels and hardware > behavior, it's very important to collect address maps with physical > addresses. For example, 3D XPoint access can only be found by filtering > the physical address. > However, perf doesn't collect physical address information in sampling. > > The load latency/DLA information in PEBS can be used to calculate the > physical address. > For kernel direct mapping addresses, virt_to_phys is used to convert the > virtual addresses from DLA to physical address. > For user virtual addresses, __get_user_pages_fast is used to walk the > pages tables for user physical address. > This does not work for vmalloc addresses. Right now these are not > resolved, but code to do that could be added. > For security, the physical address can only be exposed to root or > privileged user. > A new sample type PERF_SAMPLE_PHYS_ADDR is introduced to expose the > physical addresses. > > Signed-off-by: Kan Liang This is a useful feature that I would like to use for some of the analysis we would like to do. Please apply. I am assuming you have the support in perf record/report? > --- > > The V2 version can be found here. > https://patchwork.kernel.org/patch/7979771/ > > Changes since V2 > - Only the kernel patch > - Add example in changelog > - Include a perf_paranoid_kernel() test (PeterZ) > - Fix minor complier warning > > arch/x86/events/intel/ds.c | 24 ++++++++++++++++++++++++ > arch/x86/events/perf_event.h | 2 +- > include/linux/perf_event.h | 3 +++ > include/uapi/linux/perf_event.h | 4 +++- > kernel/events/core.c | 12 ++++++++++++ > 5 files changed, 43 insertions(+), 2 deletions(-) > > diff --git a/arch/x86/events/intel/ds.c b/arch/x86/events/intel/ds.c > index a322fed..712c86b 100644 > --- a/arch/x86/events/intel/ds.c > +++ b/arch/x86/events/intel/ds.c > @@ -1083,6 +1083,7 @@ static void setup_pebs_sample_data(struct perf_event *event, > u64 sample_type; > int fll, fst, dsrc; > int fl = event->hw.flags; > + struct page *p = NULL; > > if (pebs == NULL) > return; > @@ -1179,6 +1180,29 @@ static void setup_pebs_sample_data(struct perf_event *event, > x86_pmu.intel_cap.pebs_format >= 1) > data->addr = pebs->dla; > > + if ((sample_type & PERF_SAMPLE_PHYS_ADDR) && (data->addr != 0)) { > + if (data->addr >= TASK_SIZE) { > + /* If it's vmalloc()d memory, leave phys_addr as 0 */ > + if (virt_addr_valid(data->addr) && > + !(data->addr >= VMALLOC_START && data->addr < VMALLOC_END)) > + data->phys_addr = (u64)virt_to_phys((void *)(uintptr_t)data->addr); > + } else { > + /* > + * Walking the pages tables for user address. > + * Interrupts are disabled, so it prevents any tear down > + * of the page tables. > + * Try IRQ-safe __get_user_pages_fast first. > + * If failed, leave phys_addr as 0. > + */ > + if ((current->mm != NULL) && > + (__get_user_pages_fast(data->addr, 1, 0, &p) == 1)) > + data->phys_addr = page_to_phys(p) + data->addr % PAGE_SIZE; > + > + if (p) > + put_page(p); > + } > + } > + This would be cleaner in a separate function in my opinion. > if (x86_pmu.intel_cap.pebs_format >= 2) { > /* Only set the TSX weight when no memory weight. */ > if ((sample_type & PERF_SAMPLE_WEIGHT) && !fll) > diff --git a/arch/x86/events/perf_event.h b/arch/x86/events/perf_event.h > index 476aec3..65bb91e 100644 > --- a/arch/x86/events/perf_event.h > +++ b/arch/x86/events/perf_event.h > @@ -91,7 +91,7 @@ struct amd_nb { > (PERF_SAMPLE_IP | PERF_SAMPLE_TID | PERF_SAMPLE_ADDR | \ > PERF_SAMPLE_ID | PERF_SAMPLE_CPU | PERF_SAMPLE_STREAM_ID | \ > PERF_SAMPLE_DATA_SRC | PERF_SAMPLE_IDENTIFIER | \ > - PERF_SAMPLE_TRANSACTION) > + PERF_SAMPLE_TRANSACTION | PERF_SAMPLE_PHYS_ADDR) > > /* > * A debug store configuration. > diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h > index a3b873f..6783c69 100644 > --- a/include/linux/perf_event.h > +++ b/include/linux/perf_event.h > @@ -944,6 +944,8 @@ struct perf_sample_data { > > struct perf_regs regs_intr; > u64 stack_user_size; > + > + u64 phys_addr; > } ____cacheline_aligned; > > /* default value for data source */ > @@ -964,6 +966,7 @@ static inline void perf_sample_data_init(struct perf_sample_data *data, > data->weight = 0; > data->data_src.val = PERF_MEM_NA; > data->txn = 0; > + data->phys_addr = 0; > } > > extern void perf_output_sample(struct perf_output_handle *handle, > diff --git a/include/uapi/linux/perf_event.h b/include/uapi/linux/perf_event.h > index 642db5f..cbea02f 100644 > --- a/include/uapi/linux/perf_event.h > +++ b/include/uapi/linux/perf_event.h > @@ -139,8 +139,9 @@ enum perf_event_sample_format { > PERF_SAMPLE_IDENTIFIER = 1U << 16, > PERF_SAMPLE_TRANSACTION = 1U << 17, > PERF_SAMPLE_REGS_INTR = 1U << 18, > + PERF_SAMPLE_PHYS_ADDR = 1U << 19, > > - PERF_SAMPLE_MAX = 1U << 19, /* non-ABI */ > + PERF_SAMPLE_MAX = 1U << 20, /* non-ABI */ > }; > > /* > @@ -814,6 +815,7 @@ enum perf_event_type { > * { u64 transaction; } && PERF_SAMPLE_TRANSACTION > * { u64 abi; # enum perf_sample_regs_abi > * u64 regs[weight(mask)]; } && PERF_SAMPLE_REGS_INTR > + * { u64 phys_addr;} && PERF_SAMPLE_PHYS_ADDR > * }; > */ > PERF_RECORD_SAMPLE = 9, > diff --git a/kernel/events/core.c b/kernel/events/core.c > index 426c2ff..07d4a74 100644 > --- a/kernel/events/core.c > +++ b/kernel/events/core.c > @@ -1570,6 +1570,9 @@ static void __perf_event_header_size(struct perf_event *event, u64 sample_type) > if (sample_type & PERF_SAMPLE_TRANSACTION) > size += sizeof(data->txn); > > + if (sample_type & PERF_SAMPLE_PHYS_ADDR) > + size += sizeof(data->phys_addr); > + > event->header_size = size; > } > > @@ -5952,6 +5955,9 @@ void perf_output_sample(struct perf_output_handle *handle, > if (sample_type & PERF_SAMPLE_DATA_SRC) > perf_output_put(handle, data->data_src.val); > > + if (sample_type & PERF_SAMPLE_PHYS_ADDR) > + perf_output_put(handle, data->phys_addr); > + I think this is misplaced. You need to output in the order in which you describe the record in the header file. You have: DATA_SRC, SAMPLE_TRANSACTION, REGS_INTR, PHYS_ADDR. That means you need to output in that order. Otherwise your PHYS_ADDR may be interpreted as a SAMPLE_TRANSACTION in case it is active. At least this is my understanding of the ABI here. > if (sample_type & PERF_SAMPLE_TRANSACTION) > perf_output_put(handle, data->txn); > > @@ -9852,6 +9858,12 @@ SYSCALL_DEFINE5(perf_event_open, > return -EINVAL; > } > > + /* Only privileged users can get kernel addresses */ > + if ((attr.sample_type & PERF_SAMPLE_PHYS_ADDR) && > + perf_paranoid_kernel() && > + !capable(CAP_SYS_ADMIN)) > + return -EACCES; > + > if (!attr.sample_max_stack) > attr.sample_max_stack = sysctl_perf_event_max_stack; > > -- > 2.4.3 >