• Thomas Richter's avatar
    s390/cpum_sf: Avoid SBD overflow condition in irq handler · 0539ad0b
    Thomas Richter authored
    The s390 CPU Measurement sampling facility has an overflow condition
    which fires when all entries in a SBD are used.
    The measurement alert interrupt is triggered and reads out all samples
    in this SDB. It then tests the successor SDB, if this SBD is not full,
    the interrupt handler does not read any samples at all from this SDB
    The design waits for the hardware to fill this SBD and then trigger
    another meassurement alert interrupt.
    
    This scheme works nicely until
    an perf_event_overflow() function call discards the sample due to
    a too high sampling rate.
    The interrupt handler has logic to read out a partially filled SDB
    when the perf event overflow condition in linux common code is met.
    This causes the CPUM sampling measurement hardware and the PMU
    device driver to operate on the same SBD's trailer entry.
    This should not happen.
    
    This can be seen here using this trace:
       cpumsf_pmu_add: tear:0xb5286000
       hw_perf_event_update: sdbt 0xb5286000 full 1 over 0 flush_all:0
       hw_perf_event_update: sdbt 0xb5286008 full 0 over 0 flush_all:0
            above shows 1. interrupt
       hw_perf_event_update: sdbt 0xb5286008 full 1 over 0 flush_all:0
       hw_perf_event_update: sdbt 0xb5286008 full 0 over 0 flush_all:0
            above shows 2. interrupt
    	... this goes on fine until...
       hw_perf_event_update: sdbt 0xb5286068 full 1 over 0 flush_all:0
       perf_push_sample1: overflow
          one or more samples read from the IRQ handler are rejected by
          perf_event_overflow() and the IRQ handler advances to the next SDB
          and modifies the trailer entry of a partially filled SDB.
       hw_perf_event_update: sdbt 0xb5286070 full 0 over 0 flush_all:1
          timestamp: 14:32:52.519953
    
    Next time the IRQ handler is called for this SDB the trailer entry shows
    an overflow count of 19 missed entries.
       hw_perf_event_update: sdbt 0xb5286070 full 1 over 19 flush_all:1
          timestamp: 14:32:52.970058
    
    Remove access to a follow on SDB when event overflow happened.
    Signed-off-by: default avatarThomas Richter <tmricht@linux.ibm.com>
    Signed-off-by: default avatarVasily Gorbik <gor@linux.ibm.com>
    0539ad0b
perf_cpum_sf.c 62 KB