diff options
author | Hendrik Brueckner <brueckner@linux.vnet.ibm.com> | 2013-12-13 12:45:01 +0100 |
---|---|---|
committer | Martin Schwidefsky <schwidefsky@de.ibm.com> | 2013-12-16 14:38:01 +0100 |
commit | d7528862cf035994972c2c6f42c927db78f2f3a2 (patch) | |
tree | 8e5c54f2de4462a8600037f7563930acf5260bcf /arch/s390/kernel | |
parent | 7e75fc3ff4cffd90684816d69838f8730ac3e072 (diff) |
s390/cpum_sf: Add flag to process full SDBs only
Add the PERF_CPUM_SF_FULL_BLOCKS flag to process only sample-data-blocks that
have the block-full-indicator bit set. Sample-data-blocks that are partially
filled are discarded. Use this flag if the sampling buffer is likely to be
shared among perf events that use different sampling modes. In such
environments, flushing sample-data-blocks that are not completely filled, might
cause invalid-data-formats.
Setting PERF_CPUM_SF_FULL_BLOCKS prevents potentially invalid sampling data to
be processed but, in contrast, also discards valid samples in partially filled
sample-data-blocks. Note that sample-data-blocks might not become full for
small sampling frequencies or for workload that is scheduled for tiny intervals.
To sample with the PERF_CPUM_SF_FULL_BLOCKS flag, set the perf->attr.config1
to 0x0004. For example:
perf record -e cpum_sf/config=0xB000,config1=0x0004/
Signed-off-by: Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
Diffstat (limited to 'arch/s390/kernel')
-rw-r--r-- | arch/s390/kernel/perf_cpum_sf.c | 13 |
1 files changed, 11 insertions, 2 deletions
diff --git a/arch/s390/kernel/perf_cpum_sf.c b/arch/s390/kernel/perf_cpum_sf.c index b4ec058c4f10..3c3bc8d7b220 100644 --- a/arch/s390/kernel/perf_cpum_sf.c +++ b/arch/s390/kernel/perf_cpum_sf.c @@ -733,6 +733,10 @@ static int __hw_perf_event_init(struct perf_event *event) SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_DIAG_MODE; } + /* Check and set other sampling flags */ + if (attr->config1 & PERF_CPUM_SF_FULL_BLOCKS) + SAMPL_FLAGS(hwc) |= PERF_CPUM_SF_FULL_BLOCKS; + /* The sampling information (si) contains information about the * min/max sampling intervals and the CPU speed. So calculate the * correct sampling interval and avoid the whole period adjust @@ -1203,8 +1207,10 @@ static void hw_collect_samples(struct perf_event *event, unsigned long *sdbt, * register of the specified perf event. * * Only full sample-data-blocks are processed. Specify the flash_all flag - * to also walk through partially filled sample-data-blocks. - * + * to also walk through partially filled sample-data-blocks. It is ignored + * if PERF_CPUM_SF_FULL_BLOCKS is set. The PERF_CPUM_SF_FULL_BLOCKS flag + * enforces the processing of full sample-data-blocks only (trailer entries + * with the block-full-indicator bit set). */ static void hw_perf_event_update(struct perf_event *event, int flush_all) { @@ -1214,6 +1220,9 @@ static void hw_perf_event_update(struct perf_event *event, int flush_all) unsigned long long event_overflow, sampl_overflow, num_sdb, te_flags; int done; + if (flush_all && SDB_FULL_BLOCKS(hwc)) + flush_all = 0; + sdbt = (unsigned long *) TEAR_REG(hwc); done = event_overflow = sampl_overflow = num_sdb = 0; while (!done) { |