Re: [PATCH v1 2/3] perf session: Add swap operation for event TIME_CONV

From: Leo Yan
Date: Mon Apr 26 2021 - 03:49:20 EST


On Mon, Apr 26, 2021 at 10:26:12AM +0300, Adrian Hunter wrote:
> On 26/04/21 9:44 am, Leo Yan wrote:
> > Hi Adrian,
> >
> > On Mon, Apr 26, 2021 at 08:40:00AM +0300, Adrian Hunter wrote:
> >
> > [...]
> >
> >>> +static void perf_event__time_conv_swap(union perf_event *event,
> >>> + bool sample_id_all __maybe_unused)
> >>> +{
> >>> + size_t time_zero_size;
> >>> +
> >>> + event->time_conv.time_shift = bswap_64(event->time_conv.time_shift);
> >>> + event->time_conv.time_mult = bswap_64(event->time_conv.time_mult);
> >>> + event->time_conv.time_zero = bswap_64(event->time_conv.time_zero);
> >>> +
> >>> + time_zero_size = (void *)&event->time_conv.time_cycles - (void *)event;
> >>> + if (event->header.size > time_zero_size) {
> >>
> >> I wonder if we could have a helper for this e.g. (untested)
> >>
> >> #define event_contains(obj, mem) (obj.header.size > offsetof(typeof(obj), mem))
> >>
> >> if (event_contains(event->time_conv, time_cycles)) {
> >
> > Yeah, this is a better implementation. Will refine patch for this.
> >
> >>> + event->time_conv.time_cycles = bswap_64(event->time_conv.time_cycles);
> >>> + event->time_conv.time_mask = bswap_64(event->time_conv.time_mask);
> >>> + event->time_conv.cap_user_time_zero =
> >>> + bswap_32(event->time_conv.cap_user_time_zero);
> >>> + event->time_conv.cap_user_time_short =
> >>> + bswap_32(event->time_conv.cap_user_time_short);
> >>
> >> 'struct perf_record_time_conv' contains bool, the sizeof which, AFAIK, is not defined.
> >> Is it really 4 bytes on your implementation? It is only 1 byte with gcc on x86.
> >
> > On Arm64, bool is also 1 byte with GCC.
> >
> > When working on this patch, I checked the size for structure
> > perf_record_time_conv, it gave out the structure size is 56; I wrongly
> > thought the bool size is 4 bytes and there have no padding. In fact,
> > bool size is 1 byte and GCC pads 6 bytes at the end of structure.
> >
> >> Either way, you should change 'struct perf_record_time_conv' so it uses a type of known size.
> >> Since you are the only one using it, it should match your implementation.
> >
> > Will change to "unsigned int" value.
>
> If it is 1 byte, I'd have thought __u8 i.e.
>
> struct perf_record_time_conv {
> struct perf_event_header header;
> __u64 time_shift;
> __u64 time_mult;
> __u64 time_zero;
> __u64 time_cycles;
> __u64 time_mask;
> __u8 cap_user_time_zero;
> __u8 cap_user_time_short;
> __u8 unused[6];
> };
>

Ah, yes, thanks a lot for reminding.