Re: [PATCH] perf: cs-etm: update ETM metadata format

Hi Leo,

On Thu, 14 Jan 2021 at 13:46, Leo Yan leo.yan@linaro.org wrote:

Hi Mike,

On Wed, Jan 13, 2021 at 01:35:04PM +0000, Mike Leach wrote:

...

The current fixed metadata version format (version 0), means that adding metadata parameter items renders files from a previous version of perf unreadable. Per CPU parameters appear in a fixed order, but there is no field to indicate the number of ETM parameters per CPU.

This patch updates the per CPU parameter blocks to include a NR_PARAMs value which indicates the number of parameters in the block.

The header version is incremented to 1. Fixed ordering is retained, new ETM parameters are added to the end of the list.

The reader code is updated to be able to read current version 0 files, For version 1, the reader will read the number of parameters in the per CPU block. This allows the reader to process older or newer files that may have different numbers of parameters than in use at the time perf was built.

Seems to me, this patch is big and have several complex logics, so it's better to divide into smaller patches in one series. But this is not necessary, please see below comments.

It is quite big, but all the changes are directly related to the header file format change.

...

Signed-off-by: Mike Leach mike.leach@linaro.org

tools/perf/arch/arm/util/cs-etm.c | 3 +- tools/perf/util/cs-etm.c | 171 ++++++++++++++++++++++++------ tools/perf/util/cs-etm.h | 22 +++- 3 files changed, 159 insertions(+), 37 deletions(-)

diff --git a/tools/perf/arch/arm/util/cs-etm.c b/tools/perf/arch/arm/util/cs-etm.c index bd446aba64f7..1876cdf146a0 100644 --- a/tools/perf/arch/arm/util/cs-etm.c +++ b/tools/perf/arch/arm/util/cs-etm.c @@ -629,6 +629,7 @@ static void cs_etm_get_metadata(int cpu, u32 *offset, /* Build generic header portion */ info->priv[*offset + CS_ETM_MAGIC] = magic; info->priv[*offset + CS_ETM_CPU] = cpu;

• info->priv[*offset + CS_ETM_NR_PARAMS] = increment - CS_ETM_NR_PARAMS - 1;
  

NR_PARAMS can be a self-contained value, so it can account the items from common header, thus it could be:

info->priv[*offset + CS_ETM_NR_PARAMS] = increment;

Changing this here, will simply alter the logic elsewhere. Having NR_PARAMS as the number of specific ETM parameters makes the user readble output from --dump better.

...

  /* Where the next CPU entry should start from */
  *offset += increment;

} @@ -674,7 +675,7 @@ static int cs_etm_info_fill(struct auxtrace_record *itr,

  /* First fill out the session header */
  info->type = PERF_AUXTRACE_CS_ETM;

• info->priv[CS_HEADER_VERSION_0] = 0;
  

• info->priv[CS_HEADER_VERSION_0] = CS_HEADER_CURRENT_VERSION;
  

Should rename "CS_HEADER_VERSION_0" to "CS_HEADER_VERSION"? Also rename "CS_HEADER_VERSION_0_MAX" to "CS_HEADER_VERSION_MAX"?

Yes - could do this.

...

  info->priv[CS_PMU_TYPE_CPUS] = type << 32;
  info->priv[CS_PMU_TYPE_CPUS] |= nr_cpu;
  info->priv[CS_ETM_SNAPSHOT] = ptr->snapshot_mode;

diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c index a2a369e2fbb6..394331aa058a 100644 --- a/tools/perf/util/cs-etm.c +++ b/tools/perf/util/cs-etm.c @@ -2443,6 +2443,7 @@ static const char * const cs_etm_global_header_fmts[] = { static const char * const cs_etm_priv_fmts[] = { [CS_ETM_MAGIC] = " Magic number %llx\n", [CS_ETM_CPU] = " CPU %lld\n",

• [CS_ETM_NR_PARAMS]      = "     NR_PARAMS                      %llx\n",
  [CS_ETM_ETMCR]          = "     ETMCR                          %llx\n",
  [CS_ETM_ETMTRACEIDR]    = "     ETMTRACEIDR                    %llx\n",
  [CS_ETM_ETMCCER]        = "     ETMCCER                        %llx\n",
  

@@ -2452,6 +2453,7 @@ static const char * const cs_etm_priv_fmts[] = { static const char * const cs_etmv4_priv_fmts[] = { [CS_ETM_MAGIC] = " Magic number %llx\n", [CS_ETM_CPU] = " CPU %lld\n",

• [CS_ETM_NR_PARAMS]      = "     NR_PARAMS                      %llx\n",
  [CS_ETMV4_TRCCONFIGR]   = "     TRCCONFIGR                     %llx\n",
  [CS_ETMV4_TRCTRACEIDR]  = "     TRCTRACEIDR                    %llx\n",
  [CS_ETMV4_TRCIDR0]      = "     TRCIDR0                        %llx\n",
  

@@ -2461,24 +2463,127 @@ static const char * const cs_etmv4_priv_fmts[] = { [CS_ETMV4_TRCAUTHSTATUS] = " TRCAUTHSTATUS %llx\n", };

+static const char * const param_unk_fmt =

• "       Unknown parameter [%d]         %llx\n";
  

+static const char * const magic_unk_fmt =

• "       Magic number Unknown           %llx\n";
  

If "magic_unk_fmt" is only used once, it's not necessary to define it as static variable and can directly print it in the place.

This is here to make the print loop easier to read. The actual print statement is deeply indented, and all the other statements have format as a pre-defined string.

...

static void cs_etm__print_auxtrace_info(__u64 *val, int num) {

• int i, j, cpu = 0;
  

• int i, j, cpu = 0, version, nr_params;
  

• __u64 magic;
  
  

• fprintf(stdout, "       Reader Header version          %llx\n",
  

•         (__u64)CS_HEADER_CURRENT_VERSION);
  

This is redundant, the version will be printed by below loop.

No it is not. If the new version perf reads a version 0 file then the following is output:

0x188 [0x148]: PERF_RECORD_AUXTRACE_INFO type: 3 Reader Header version 1 Header version 0 PMU type/num cpus 900000004

...

  for (i = 0; i < CS_HEADER_VERSION_0_MAX; i++)

...

          fprintf(stdout, cs_etm_global_header_fmts[i], val[i]);

• version = val[0];
  
  for (i = CS_HEADER_VERSION_0_MAX; cpu < num; cpu++) {
  

Considering in the same loop, it needs to handle version 0 and 1 saperately, so it's better to use two functions to encapsulate metadata dumping:

    int version = val[0];

    for (i = CS_HEADER_VERSION_0_MAX; cpu < num; cpu++) {
            if (version == 0)
                    cs_etm__print_metadata_v0(cpu, val, i);
            else if (version == 1)
                    cs_etm__print_metadata_v1(cpu, val, i);
            else /* Not supported */
                    return;
    }

OK - I'll see if this can be made feasible.

...

•         if (val[i] == __perf_cs_etmv3_magic)
  

•                 for (j = 0; j < CS_ETM_PRIV_MAX; j++, i++)
  

•                         fprintf(stdout, cs_etm_priv_fmts[j], val[i]);
  

•         else if (val[i] == __perf_cs_etmv4_magic)
  

•                 for (j = 0; j < CS_ETMV4_PRIV_MAX; j++, i++)
  

•                         fprintf(stdout, cs_etmv4_priv_fmts[j], val[i]);
  

•         else
  

•         /* common header block */
  

•         magic = val[i];
  

•         fprintf(stdout, cs_etm_priv_fmts[0], val[i++]);
  

•         fprintf(stdout, cs_etm_priv_fmts[1], val[i++]);
  

•         /* nr param only present in version 1 */
  

•         if (version) {
  

•                 nr_params = val[i];
  

•                 fprintf(stdout, cs_etm_priv_fmts[2], val[i++]);
  

•         }
  

•         if (magic == __perf_cs_etmv3_magic) {
  

•                 /* version 0 - fixed number of params */
  

•                 if (!version)
  

•                         nr_params =  CS_ETM_NR_PARAMS_V0;
  

Redundant space after "=".

OK

...

•                 /* after common block */
  

•                 for (j = 3; j < nr_params + 3; j++, i++) {
  

This is not correct, for the perf data with version 0, you should start to print from item 2 for CS_ETM_ETMCR; Only for version 1, it should start from 3 for CS_ETM_ETMCR.

The 'j' indexes into the print formats which are v1 ordered, not the incoming data. So we have to skip the index to NR_PARAMS when printing v0 files. But worth a comment here I think.

...

•                         /* if newer record - could be excess params */
  

•                         if (j >= CS_ETM_PRIV_MAX)
  

I think we need support backward compatibility, but cannot support forward compatibility. So need to report error for this?

This depends on the content of the trace data. The discussions we had within ARM were triggered when a perf.data file from a customer who had added their own parameter to the list failed the perf report --dump when we tried to analyse it. The idea of forward compatibily is that you can at least see there are additional parameters, which makes support tasks a little easier.

If a newer version of perf is used in this sort of case to obtain data and pass it on, analysis is still possible if the new parameter is not directly relevant for the decode for a particular test. But if analysis does fail - then perf report --dump will show the "Unknown" parameters in the file, rather than silently crash out as the case is now.

...

•                                 fprintf(stdout, param_unk_fmt, j, val[i]);
  

•                         else
  

•                                 fprintf(stdout, cs_etm_priv_fmts[j], val[i]);
  

•                 }
  

•         } else if (magic == __perf_cs_etmv4_magic) {
  

•                 /* version 0 - fixed number of params */
  

•                 if (!version)
  

•                         nr_params = CS_ETMV4_NR_PARAMS_V0;
  

•                 /* after common block */
  

•                 for (j = 3; j < nr_params + 3; j++, i++) {
  

•                         /* if newer record - could be excess params */
  

•                         if (j >= CS_ETMV4_PRIV_MAX)
  

•                                 fprintf(stdout, param_unk_fmt, j, val[i]);
  

•                         else
  

•                                 fprintf(stdout, cs_etmv4_priv_fmts[j], val[i]);
  

•                 }
  

•         } else {
                  /* failure.. return */
  

•                 fprintf(stdout, magic_unk_fmt, magic);
                  return;
  

•         }
  

• }
  

+}

+/*

• • Read a single cpu parameter block from the auxtrace_info priv block.
• • For version 1 there is a per cpu nr_params entry. If we are handling
• • version 1 file, then there may be less, the same, or more params
• • indicated by this value than the compile time number we understand.
• • For a version 0 info block, there are a fixed number, and we need to
• • fill out the nr_param value in the metadata we create.
• */

+static u64 *cs_etm__create_meta_blk(u64 *buff_in, int *buff_in_offset,

•                             int out_blk_size, int nr_params_v0)
  

+{

• u64 *metadata = NULL;
  

• int hdr_version;
  

• int nr_in_params, nr_out_params, nr_cmn_params;
  

• int i, k;
  

• metadata = zalloc(sizeof(*metadata) * out_blk_size);
  

• if (!metadata)
  

•         return NULL;
  

• /* read block current index & version */
  

• i = *buff_in_offset;
  

• hdr_version = buff_in[CS_HEADER_VERSION_0];
  

• if (!hdr_version) {
  

• /* read version 0 info block into a version 1 metadata block  */
  

•         nr_in_params = nr_params_v0;
  

•         metadata[CS_ETM_MAGIC] = buff_in[i + CS_ETM_MAGIC];
  

•         metadata[CS_ETM_CPU] = buff_in[i + CS_ETM_CPU];
  

•         metadata[CS_ETM_NR_PARAMS] = nr_in_params;
  

•         /* remaining block params at offset +1 from source */
  

•         for (k = 2; k < nr_in_params; k++)
  

•                 metadata[k+1] = buff_in[i + k];
  

•         /* version 0 has 2 common params */
  

•         nr_cmn_params = 2;
  

• } else {
  

• /* read version 1 info block - input and output nr_params may differ */
  

•         /* version 1 has 3 common params */
  

•         nr_cmn_params = 3;
  

•         nr_in_params = buff_in[i + CS_ETM_NR_PARAMS];
  

•         /* if input has more params than output - skip excess */
  

•         nr_out_params = nr_in_params + nr_cmn_params;
  

•         if (nr_out_params > out_blk_size)
  

•                 nr_out_params = out_blk_size;
  

•         for (k = 0; k < nr_out_params; k++)
  

•                 metadata[k] = buff_in[i + k];
  

•         /* record the actual nr params we copied */
  

•         metadata[CS_ETM_NR_PARAMS] = nr_out_params - nr_cmn_params;
  

If we set CS_ETM_NR_PARAMS as self-contained value (it presents all item numbers from CS_ETM_MAGIC to CS_ETMV4_TRCAUTHSTATUS), and we can just support backward compatibility and don't support forward compabitility, I think the logic can be simplized.

See my comments above for the reason we want to support limited forward compatibility.

Thanks for the review.

Regards

Mike

Thanks, Leo

...

  }

• /* adjust in offset by number of in params used */
  

• i += nr_in_params + nr_cmn_params;
  

• *buff_in_offset = i;
  

• return metadata;
  

}

int cs_etm__process_auxtrace_info(union perf_event *event, @@ -2492,11 +2597,12 @@ int cs_etm__process_auxtrace_info(union perf_event *event, int info_header_size; int total_size = auxtrace_info->header.size; int priv_size = 0;

• int num_cpu;
  

• int err = 0, idx = -1;
  

• int i, j, k;
  

• int num_cpu, trcidr_idx;
  

• int err = 0;
  

• int i, j;
  u64 *ptr, *hdr = NULL;
  u64 **metadata = NULL;
  

• u64 hdr_version;
  
  /*
   * sizeof(auxtrace_info_event::type) +
  

@@ -2512,8 +2618,9 @@ int cs_etm__process_auxtrace_info(union perf_event *event, /* First the global part */ ptr = (u64 *) auxtrace_info->priv;

• /* Look for version '0' of the header */
  

• if (ptr[0] != 0)
  

• /* Look for version of the header */
  

• hdr_version = ptr[0];
  

• if (hdr_version > CS_HEADER_CURRENT_VERSION)
          return -EINVAL;
  
  hdr = zalloc(sizeof(*hdr) * CS_HEADER_VERSION_0_MAX);
  

@@ -2552,35 +2659,31 @@ int cs_etm__process_auxtrace_info(union perf_event *event, */ for (j = 0; j < num_cpu; j++) { if (ptr[i] == __perf_cs_etmv3_magic) {

•                 metadata[j] = zalloc(sizeof(*metadata[j]) *
  

•                                      CS_ETM_PRIV_MAX);
  

•                 if (!metadata[j]) {
  

•                         err = -ENOMEM;
  

•                         goto err_free_metadata;
  

•                 }
  

•                 for (k = 0; k < CS_ETM_PRIV_MAX; k++)
  

•                         metadata[j][k] = ptr[i + k];
  

•                 metadata[j] =
  

•                         cs_etm__create_meta_blk(ptr, &i,
  

•                                                 CS_ETM_PRIV_MAX,
  

•                                                 CS_ETM_NR_PARAMS_V0);
  
                  /* The traceID is our handle */
  

•                 idx = metadata[j][CS_ETM_ETMTRACEIDR];
  

•                 i += CS_ETM_PRIV_MAX;
  

•                 trcidr_idx = CS_ETM_ETMTRACEIDR;
  

•         } else if (ptr[i] == __perf_cs_etmv4_magic) {
  

•                 metadata[j] = zalloc(sizeof(*metadata[j]) *
  

•                                      CS_ETMV4_PRIV_MAX);
  

•                 if (!metadata[j]) {
  

•                         err = -ENOMEM;
  

•                         goto err_free_metadata;
  

•                 }
  

•                 for (k = 0; k < CS_ETMV4_PRIV_MAX; k++)
  

•                         metadata[j][k] = ptr[i + k];
  

•                 metadata[j] =
  

•                         cs_etm__create_meta_blk(ptr, &i,
  

•                                                 CS_ETMV4_PRIV_MAX,
  

•                                                 CS_ETMV4_NR_PARAMS_V0);
  
                  /* The traceID is our handle */
  

•                 idx = metadata[j][CS_ETMV4_TRCTRACEIDR];
  

•                 i += CS_ETMV4_PRIV_MAX;
  

•                 trcidr_idx = CS_ETMV4_TRCTRACEIDR;
  

•         }
  

•         if (!metadata[j]) {
  

•                 err = -ENOMEM;
  

•                 goto err_free_metadata;
          }
  
          /* Get an RB node for this CPU */
  

•         inode = intlist__findnew(traceid_list, idx);
  

•         inode = intlist__findnew(traceid_list, metadata[j][trcidr_idx]);
  
          /* Something went wrong, no need to continue */
          if (!inode) {
  

diff --git a/tools/perf/util/cs-etm.h b/tools/perf/util/cs-etm.h index 4ad925d6d799..fffc6b50beb9 100644 --- a/tools/perf/util/cs-etm.h +++ b/tools/perf/util/cs-etm.h @@ -24,16 +24,28 @@ enum { CS_HEADER_VERSION_0_MAX, };

+/*

• • Update the version for new format.
• • New version 1 format adds a param count to the per cpu metadata.
• • This allows easy adding of new metadata parameters.
• • Requires that new params always added after current ones.
• • Also allows client reader to handle file versions that are different by
• • checking the number of params in the file vs the number expected.
• */

+#define CS_HEADER_CURRENT_VERSION 1

/* Beginning of header common to both ETMv3 and V4 */ enum { CS_ETM_MAGIC, CS_ETM_CPU,

• CS_ETM_NR_PARAMS,
  

};

/* ETMv3/PTM metadata */ enum { /* Dynamic, configurable parameters */

• CS_ETM_ETMCR = CS_ETM_CPU + 1,
  

• CS_ETM_ETMCR = CS_ETM_NR_PARAMS + 1,
  CS_ETM_ETMTRACEIDR,
  /* RO, taken from sysFS */
  CS_ETM_ETMCCER,
  

@@ -41,10 +53,13 @@ enum { CS_ETM_PRIV_MAX, };

+/* define fixed version 0 length - allow new format reader to read old files. */ +#define CS_ETM_NR_PARAMS_V0 (CS_ETM_ETMIDR - CS_ETM_ETMCR + 1)

/* ETMv4 metadata */ enum { /* Dynamic, configurable parameters */

• CS_ETMV4_TRCCONFIGR = CS_ETM_CPU + 1,
  

• CS_ETMV4_TRCCONFIGR = CS_ETM_NR_PARAMS + 1,
  CS_ETMV4_TRCTRACEIDR,
  /* RO, taken from sysFS */
  CS_ETMV4_TRCIDR0,
  

@@ -55,6 +70,9 @@ enum { CS_ETMV4_PRIV_MAX, };

+/* define fixed version 0 length - allow new format reader to read old files. */ +#define CS_ETMV4_NR_PARAMS_V0 (CS_ETMV4_TRCAUTHSTATUS - CS_ETMV4_TRCCONFIGR + 1)

/*

• ETMv3 exception encoding number:
• See Embedded Trace Macrocell spcification (ARM IHI 0014Q)

-- 2.17.1

-- Mike Leach Principal Engineer, ARM Ltd. Manchester Design Centre. UK