Add a warning against having multiple copies of a shared library loaded

[petvana]

This PR aims to solve #42896 by reincarnation of #7490 originally created by @vtjnash.

Since PyCall loads libraries on the fly, it also need to be updated as well (see this PR). That is also the reason why check_dllist is exported.

[ararslan]

IMO it would make more sense for this to be an optional, opt-in safety check built into dlopen rather than a separate function.

[petvana]

@ararslan Agree that the warning should be optional. However, this PR targets beginner users accidentally using two incompatible packages (like Gtk and PyPlot). Thus, IMHO, I'd enable it by default (opt-out), like Bounds checking in Julia.

It is not necessary to export new function, we can duplicate the code to PyCall for the case new libraries are loaded by pyimport and not by dlopen.

We can also suppress the warning if two exactly same files but from different locations are loaded (symlinks are already detected).

[petvana]

CI produces the following two warnings.

For macos64

┌ Warning: Detected possible duplicate library loaded: libunwind
│ This may lead to unexpected behavior!
│ /usr/lib/system/libunwind.dylib
│ /Users/julia/buildbot/worker-tabularasa/tester_macos64/build/lib/julia/libunwind.1.0.dylib
└ @ Base.Libc.Libdl libdl.jl:114

For freebsd64

┌ Warning: Detected possible duplicate library loaded: libatomic
│ This may lead to unexpected behavior!
│ /usr/home/julia/buildbot/w2_tester/tester_freebsd64/build/lib/julia/libatomic.so.3
│ /usr/local/lib/gcc10/libatomic.so.1
└ @ Base.Libc.Libdl libdl.jl:114

Is it correct, or can be harmful? I cannot reproduce locally because I use Ubuntu.

[giordano]

Have you measured the performance impact of this? When loading certain packages we may dlopen some dozens of libraries, even if a single check may not be too expensive, it all adds up, and you're going to iterate every time on a larger dllist

[petvana]

@giordano Excellent point. Right now, it seems negligible for average packages. However, the complexity is right now O(n n^2) = O(n^3) of string comparisons, where n is the number of loaded libs. By sorting, we can easily go down to O(n^2 log n). Do you have in mind an example of a package that loads a lot of libraries to be tested? In such a case, the specific package can opt-out and test duplicity only once when initialized.

[petvana]

EDITED: I have improved the performance by sorting a dictionary. So, now the complexity is O(n^2 log n) O(n^2). The time (of a single check) with 215 libraries was reduced from about 1.5 ms to 263.650 μs 39 μs. (Updated after Add comments and note to docs commit)

julia> using Gtk

julia> using PyPlot
┌ Warning: Detected possible duplicate library loaded: libffi
│ This may lead to unexpected behavior!
│ /home/petr/.julia/conda/3/lib/python3.7/lib-dynload/../../libffi.so.7
│ /home/petr/.julia/artifacts/f1a54fe617cfa5997fa86c2cb15b93ac1d8c4c24/lib64/libffi.so
│ To suppress this warning, you can use the following argument:
│ dlopen([name_or_path]; suppress_warnings = ["libffi"])
└ @ Base.Libc.Libdl libdl.jl:116
┌ Warning: Detected possible duplicate library loaded: libjpeg
│ This may lead to unexpected behavior!
│ /home/petr/.julia/conda/3/lib/python3.7/site-packages/PIL/../../../libjpeg.so.9
│ /home/petr/.julia/artifacts/b9b138178f2c19d38aa8a19eac1c256f54bb6234/lib/libjpeg.so
│ To suppress this warning, you can use the following argument:
│ dlopen([name_or_path]; suppress_warnings = ["libjpeg"])
└ @ Base.Libc.Libdl libdl.jl:116

julia> using BenchmarkTools

julia> using Libdl

julia> Libdl.dllist() |> length
214

julia> @btime dllist();
  14.440 μs (434 allocations: 39.23 KiB)

julia> @btime dlopen("")
  38.515 μs (449 allocations: 64.68 KiB)

[giordano]

Do you have in mind an example of a package that loads a lot of libraries to be tested?

Gtk.jl dlopens exactly 100 libraries on Linux according to my estimation. We've spent lot of time bringing the loading time of GTK3_jll under 10 ms, this PR would kill that.

In such a case, the specific package can opt-out and test duplicity only once when initialized.

That isn't really feasible. Libraries are dlopened by JLL packages, not by a single package.

[KristofferC]

Just time loading Gtk_jll before and after this PR to get the actual numbers?

[petvana]

Tested on Intel(R) Core(TM) i5-1035G1 CPU @ 1.00GHz (Updated after the Make PR thread-safe commit)
Master:

julia> @time using GTK3_jll
  0.108701 seconds (498.00 k allocations: 21.129 MiB, 26.65% compilation time)

PR:

julia> @time using GTK3_jll
  0.110910 seconds (527.88 k allocations: 23.245 MiB, 21.68% compilation time)

Tested on Intel(R) Core(TM) i7-9700 CPU @ 3.00GHz
Master:

julia> @time using GTK3_jll
  0.081210 seconds (474.67 k allocations: 20.426 MiB, 23.50% compilation time)

PR:

julia> @time using GTK3_jll
  0.083450 seconds (505.28 k allocations: 22.627 MiB, 20.80% compilation time)

(Most of the time is spend in dlname() calling basename() that can be optimized.)

[giordano]

This is not thread-safe and goes against #41602

[petvana]

Thank you for the comment; it should be thread-safe now (without a performance regression).

[petvana]

Function check_dllist() is not exported any more, and one can execute the check by Libdl.dlopen(""), like in JuliaPy/PyCall.jl#942

[petvana]

@giordano I've added an option to suppress any future warnings for a specific library, as the warning informs. Would it work for libpetsc?

julia> using PyCall

julia> using GTK3_jll
┌ Warning: Detected possible duplicate library loaded: libffi
│ This may lead to unexpected behavior!
│ /home/petr/.julia/artifacts/f1a54fe617cfa5997fa86c2cb15b93ac1d8c4c24/lib64/libffi.so
│ /home/petr/.julia/conda/3/lib/python3.8/lib-dynload/../../libffi.so.7
│ To suppress this warning, you can use the following argument:
│ dlopen([name_or_path]; suppress_warnings = ["libffi"])
└ @ Base.Libc.Libdl libdl.jl:110

[petvana]

@vtjnash Would you please check this PR (as the author of the original idea)? I've improved performance to spend only about 2-3% when loading GTK3_jll (one of the worst cases).

[JeffBezanson]

Triage is ok with this, as long as users won't be spammed with too many or duplicate warnings. Performance impact also needs to be negligible, which it looks like it is.

[petvana]

Wow, the triaging was fast, thanks.

First, it would be convenient to have Julia itself warning-free. The CI already found at least two possible duplicities (for macos64 and freebsd64 as mentioned above #43007 (comment)). Is it expected behavior? (Unfortunately, I don't use any of these OSes now.)

[vtjnash]

triage noted we need to solve those existing warnings first

[JeffBezanson]

Yeah, if there are problems there we should fix them first, and also need to make sure the warning is 100% accurate or it will be a bad user experience. We don't want to train people to ignore warnings.