From: Thomas Richter tmricht@linux.ibm.com
[ Upstream commit 4141b6a5e9f171325effc36a22eb92bf961e7a5c ]
When perf record -e SF_CYCLES_BASIC_DIAG runs with very high frequency, the samples arrive faster than the perf process can save them to file. Eventually, for longer running processes, this leads to the siutation where the trace buffers allocated by perf slowly fills up. At one point the auxiliary trace buffer is full and the CPU Measurement sampling facility is turned off. Furthermore a warning is printed to the kernel log buffer:
cpum_sf: The AUX buffer with 0 pages for the diagnostic-sampling mode is full
The number of allocated pages for the auxiliary trace buffer is shown as zero pages. That is wrong.
Fix this by saving the number of allocated pages before entering the work loop in the interrupt handler. When the interrupt handler processes the samples, it may detect the buffer full condition and stop sampling, reducing the buffer size to zero. Print the correct value in the error message:
cpum_sf: The AUX buffer with 256 pages for the diagnostic-sampling mode is full
Signed-off-by: Thomas Richter tmricht@linux.ibm.com Signed-off-by: Vasily Gorbik gor@linux.ibm.com Signed-off-by: Sasha Levin sashal@kernel.org --- arch/s390/kernel/perf_cpum_sf.c | 1 + 1 file changed, 1 insertion(+)
diff --git a/arch/s390/kernel/perf_cpum_sf.c b/arch/s390/kernel/perf_cpum_sf.c index 5bfb1ce..74a296c 100644 --- a/arch/s390/kernel/perf_cpum_sf.c +++ b/arch/s390/kernel/perf_cpum_sf.c @@ -1537,6 +1537,7 @@ static void hw_collect_aux(struct cpu_hw_sf *cpuhw) perf_aux_output_end(handle, size); num_sdb = aux->sfb.num_sdb;
+ num_sdb = aux->sfb.num_sdb; while (!done) { /* Get an output handle */ aux = perf_aux_output_begin(handle, cpuhw->event);