On Thu, 27 Nov 2025 at 16:57, Ryan Roberts <[email protected]> wrote: > > On 27/11/2025 15:03, Ard Biesheuvel wrote: > > On Thu, 27 Nov 2025 at 15:18, Ryan Roberts <[email protected]> wrote: > >> > >> On 27/11/2025 12:28, Ard Biesheuvel wrote: > >>> On Thu, 27 Nov 2025 at 13:12, Ryan Roberts <[email protected]> wrote: > >>>> > >>>> On 27/11/2025 09:22, Ard Biesheuvel wrote: > >>>>> From: Ard Biesheuvel <[email protected]> > >>>>> > >>>>> Ryan reports that get_random_u16() is dominant in the performance > >>>>> profiling of syscall entry when kstack randomization is enabled [0]. > >>>>> > >>>>> This is the reason many architectures rely on a counter instead, and > >>>>> that, in turn, is the reason for the convoluted way the (pseudo-)entropy > >>>>> is gathered and recorded in a per-CPU variable. > >>>>> > >>>>> Let's try to make the get_random_uXX() fast path faster, and switch to > >>>>> get_random_u8() so that we'll hit the slow path 2x less often. Then, > >>>>> wire it up in the syscall entry path, replacing the per-CPU variable, > >>>>> making the logic at syscall exit redundant. > >>>> > >>>> I ran the same set of syscall benchmarks for this series as I've done > >>>> for my > >>>> series. > >>>> > >>> > >>> Thanks! > >>> > >>> > >>>> The baseline is v6.18-rc5 with stack randomization turned *off*. So I'm > >>>> showing > >>>> performance cost of turning it on without any changes to the > >>>> implementation, > >>>> then the reduced performance cost of turning it on with my changes > >>>> applied, and > >>>> finally cost of turning it on with Ard's changes applied: > >>>> > >>>> arm64 (AWS Graviton3): > >>>> +-----------------+--------------+-------------+---------------+-----------------+ > >>>> | Benchmark | Result Class | v6.18-rc5 | per-task-prng | > >>>> fast-get-random | > >>>> | | | rndstack-on | | > >>>> | > >>>> +=================+==============+=============+===============+=================+ > >>>> | syscall/getpid | mean (ns) | (R) 15.62% | (R) 3.43% | > >>>> (R) 11.93% | > >>>> | | p99 (ns) | (R) 155.01% | (R) 3.20% | > >>>> (R) 11.00% | > >>>> | | p99.9 (ns) | (R) 156.71% | (R) 2.93% | > >>>> (R) 11.39% | > >>>> +-----------------+--------------+-------------+---------------+-----------------+ > >>>> | syscall/getppid | mean (ns) | (R) 14.09% | (R) 2.12% | > >>>> (R) 10.44% | > >>>> | | p99 (ns) | (R) 152.81% | 1.55% | > >>>> (R) 9.94% | > >>>> | | p99.9 (ns) | (R) 153.67% | 1.77% | > >>>> (R) 9.83% | > >>>> +-----------------+--------------+-------------+---------------+-----------------+ > >>>> | syscall/invalid | mean (ns) | (R) 13.89% | (R) 3.32% | > >>>> (R) 10.39% | > >>>> | | p99 (ns) | (R) 165.82% | (R) 3.51% | > >>>> (R) 10.72% | > >>>> | | p99.9 (ns) | (R) 168.83% | (R) 3.77% | > >>>> (R) 11.03% | > >>>> +-----------------+--------------+-------------+---------------+-----------------+ > >>>> > >>> > >>> What does the (R) mean? > >>> > >>>> So this fixes the tail problem. I guess get_random_u8() only takes the > >>>> slow path > >>>> every 768 calls, whereas get_random_u16() took it every 384 calls. I'm > >>>> not sure > >>>> that fully explains it though. > >>>> > >>>> But it's still a 10% cost on average. > >>>> > >>>> Personally I think 10% syscall cost is too much to pay for 6 bits of > >>>> stack > >>>> randomisation. 3% is better, but still higher than we would all prefer, > >>>> I'm sure. > >>>> > >>> > >>> Interesting! > >>> > >>> So the only thing that get_random_u8() does that could explain the > >>> delta is calling into the scheduler on preempt_enable(), given that it > >>> does very little beyond that. > >>> > >>> Would you mind repeating this experiment after changing the > >>> put_cpu_var() to preempt_enable_no_resched(), to test this theory? > >> > >> This has no impact on performance. > >> > > > > Thanks. But this is really rather surprising: what else could be > > taking up that time, given that on the fast path, there are only some > > loads and stores to the buffer, and a cmpxchg64_local(). Could it be > > the latter that is causing so much latency? I suppose the local > > cmpxchg() semantics don't really exist on arm64, and this uses the > > exact same LSE instruction that would be used for an ordinary > > cmpxchg(), unlike on x86 where it appears to omit the LOCK prefix. > > > > In any case, there is no debate that your code is faster on arm64. > > The results I have for x86 show it's faster than the rdtsc too, although > that's > also somewhat surprising. I'll run your series on x86 to get the equivalent > data. >
OK, brown paper bag time ... I swapped the order of the 'old' and 'new' cmpxchg64_local() arguments, resulting in some very odd behavior. I think this explains why the tail latency was eliminated entirely, which is bizarre. The speedup is also more modest now (~2x), which may still be worthwhile, but likely insufficient for the kstack randomization case. https://git.kernel.org/pub/scm/linux/kernel/git/ardb/linux.git/log/?h=lockless-random-v2
