Re: [linus:master] [locking] c8afaa1b0f: stress-ng.zero.ops_per_sec 6.3% improvement

On 8/15/23, Linus Torvalds <[email protected]> wrote:
> On Tue, 15 Aug 2023 at 07:12, kernel test robot <[email protected]>
> wrote:
>>
>> kernel test robot noticed a 6.3% improvement of stress-ng.zero.ops_per_sec
>> on:
>
> WTF? That's ridiculous. Why would that even test new_inode() at all?
> And why would it make any difference anyway to prefetch a new inode?
> The 'zero' test claims to just read /dev/zero in a loop...
>
> [ Goes looking ]
>

Ye man, I was puzzled myself but just figured it out and was about to respond ;)

# bpftrace -e 'kprobe:new_inode { @[kstack()] = count(); }'
Attaching 1 probe...

@[
    new_inode+1
    shmem_get_inode+137
    __shmem_file_setup+195
    shmem_zero_setup+46
    mmap_region+1937
    do_mmap+956
    vm_mmap_pgoff+224
    do_syscall_64+46
    entry_SYSCALL_64_after_hwframe+115
]: 2689570

the bench is doing this *A LOT* and this looks so fishy, for the bench
itself and the kernel doing it, but I'm not going to dig into any of
that.

>>      39.35            -0.3       39.09
>> perf-profile.calltrace.cycles-pp.inode_sb_list_add.new_inode.shmem_get_inode.__shmem_file_setup.shmem_zero_setup
>
> Ahh. It also does the mmap side, and the shared case ends up always
> creating a new inode.
>
> And while the test only tests *reading* and the mmap is read-only, the
> /dev/zero file descriptor was opened for writing too, for a different
> part of a test.
>
> So even though the mapping is never written to, MAYWRITE is set, and
> so the /dev/zero mapping is done as a shared memory mapping and we
> can't do it as just a private one.
>
> That's kind of stupid and looks unintentional, but whatever.
>
> End result: that benchmark ends up being at least partly (and a fairly
> noticeable part) a shmem setup benchmark, for no actual good reason.
>
> Oh well. I certainly don't mind the removal apparently then also
> helping some odd benchmark case, but I don't think this translates to
> anything real. Very random.
>
>                     Linus
>


-- 
Mateusz Guzik <mjguzik gmail.com>