Re: [PATCH V4] perf: Add PERF_SAMPLE_PHYS_ADDR

[Stephane Eranian]

On Thu, Aug 17, 2017 at 8:51 AM,  <kan.liang@xxxxxxxxx> wrote:
> From: Kan Liang <kan.liang@xxxxxxxxx>
>
> 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 <kan.liang@xxxxxxxxx>
> ---
>
> This patch is kernel patch.
> The user space patch will be sent out later separately.
>
> Changes since V3
>  - Move the code dla->phys to separate function (Stephane)
>  - Correct PHYS_ADDR misplacement in header file (Stephane)
>
> 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      | 32 ++++++++++++++++++++++++++++++++
>  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, 51 insertions(+), 2 deletions(-)
>
> diff --git a/arch/x86/events/intel/ds.c b/arch/x86/events/intel/ds.c
> index a322fed..f0e8d9c 100644
> --- a/arch/x86/events/intel/ds.c
> +++ b/arch/x86/events/intel/ds.c
> @@ -1065,6 +1065,35 @@ static inline u64 intel_hsw_transaction(struct pebs_record_skl *pebs)
>         return txn;
>  }
>
> +static u64 dla_to_phys(u64 dla)
> +{
> +       u64 phys_addr = 0;
> +       struct page *p = NULL;
> +
> +       if (dla >= TASK_SIZE) {
> +               /* If it's vmalloc()d memory, leave phys_addr as 0 */
> +               if (virt_addr_valid(dla) &&
> +                   !(dla >= VMALLOC_START && dla < VMALLOC_END))
> +                       phys_addr = (u64)virt_to_phys((void *)(uintptr_t)dla);
> +       } 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(dla, 1, 0, &p) == 1))
> +                       phys_addr = page_to_phys(p) + dla % PAGE_SIZE;
> +
> +               if (p)
> +                       put_page(p);
> +       }
> +
> +       return phys_addr;
> +}
> +
>  static void setup_pebs_sample_data(struct perf_event *event,
>                                    struct pt_regs *iregs, void *__pebs,
>                                    struct perf_sample_data *data,
> @@ -1179,6 +1208,9 @@ 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))
> +               data->phys_addr = dla_to_phys(data->addr);
> +
>         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..c842498 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;
>  }
>
> @@ -5955,6 +5958,9 @@ void perf_output_sample(struct perf_output_handle *handle,
>         if (sample_type & PERF_SAMPLE_TRANSACTION)
>                 perf_output_put(handle, data->txn);
>
> +       if (sample_type & PERF_SAMPLE_PHYS_ADDR)
> +               perf_output_put(handle, data->phys_addr);
> +
No. That would be breaking backward compatibility with measurements
using SAMPLE_TRANSACTION and REGS_INTR. They would expect
the two fields to be laid out next to each other in the buffer.
SAMPLE_PHYS has to be the last one pushed out just like you correctly indicated
in the header file.

>         if (sample_type & PERF_SAMPLE_REGS_INTR) {
>                 u64 abi = data->regs_intr.abi;
>                 /*
> @@ -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
>