Duplicate procfs file pointer instead of reopening

[PeakKS]

This is another significant speedup to GetPageBits. By only opening the file once statically, and just dup-ing it to a local working file pointer it should stay safe while removing most fopen overhead. Benchmarks:

Elapsed CPU Time (CURRENT) = '4.29449' sec.
Elapsed CPU Time per Iteration (CURRENT, 500000) = '8.588978e-06' sec.
Elapsed CPU Time (DUP) = '0.291668' sec.
Elapsed CPU Time per Iteration (DUP, 500000) = '5.833360e-07' sec.

[KyleSanderson]

This should likely go in a RAII container like dvander mentioned on the previous PR - that way the resource are properly released on unload.

[PeakKS]

This should likely go in a RAII container like dvander mentioned on the previous PR - that way the resource are properly released on unload.

I'm no C++ whiz but I think this is the right way to do that:

static std::unique_ptr<FILE, int(*)(FILE*)> pMasterFD(fopen("/proc/self/maps", "r"), fclose);
std::unique_ptr<FILE, int(*)(FILE*)> pF(fdopen(dup(fileno(pMasterFD.get())), "r"), fclose);

That works though it is slightly slower.

Elapsed CPU Time (CURRENT) = '4.22251' sec.
Elapsed CPU Time per Iteration (CURRENT, 500000) = '8.445022e-06' sec.
Elapsed CPU Time (DUP) = '0.289283' sec.
Elapsed CPU Time per Iteration (DUP, 500000) = '5.785660e-07' sec.
Elapsed CPU Time (UNIQUE) = '0.32418' sec.
Elapsed CPU Time per Iteration (UNIQUE, 500000) = '6.483600e-07' sec.

If only the static one is made unique_ptr the time penalty is very small.

Elapsed CPU Time (CURRENT) = '4.26076' sec.
Elapsed CPU Time per Iteration (CURRENT, 500000) = '8.521512e-06' sec.
Elapsed CPU Time (DUP) = '0.28382' sec.
Elapsed CPU Time per Iteration (DUP, 500000) = '5.676400e-07' sec.
Elapsed CPU Time (UNIQUE) = '0.289645' sec.
Elapsed CPU Time per Iteration (UNIQUE, 500000) = '5.792900e-07' sec.

I'm thinking leave the working pointer as normal but change the static to unique_ptr so it is sure to be cleaned properly.

[sapphonie]

I almost wonder if we want to wait and see if the c++17 change gets merged, and then use std::filesystem ?

[sapphonie]

I almost wonder if we want to wait and see if the c++17 change gets merged, and then use std::filesystem ?

Plus, i think it'd probably also be better to do this in a ctor instead of statically in SH

[PeakKS]

Did a bit more testing but it seems the best impl is still making the static master a unique_ptr and duping it. Should be all good now.

[PeakKS]

Nevermind with further testing I have found out that the position is inherent to the file descriptor and so dup does not solve this. A thread_local unique_ptr with fseek is a little faster at least...

Elapsed CPU Time (CURRENT) = '4.53125' sec.
Elapsed CPU Time per Iteration (CURRENT, 500000) = '9.062500e-06' sec.
Elapsed CPU Time (THREAD) = '3.86515' sec.
Elapsed CPU Time per Iteration (THREAD, 500000) = '7.730302e-06' sec.

But I'd like to find something better.

[PeakKS]

Closing due to the initial premise being broken. There’s some other ideas for optimization but they should be their own PRs.