Windows test failures in file.jl: "stat: operation not permitted (EPERM)"

[tkelman]

This has been seen before by @magistere here. There may still be file handles open so the unlinked file and removed directory cannot be stat-ed. Very similar problem occurs in node: nodejs/node-v0.x-archive#7164. I tried to clean up as much as I could figure out:

ccall(:fclose, Void, (Ptr{Void},), FILEp.ptr)
ccall(:_close, Void, (Ptr{Void},), FILEp.ptr)
FILEp = nothing
a_stat = nothing
af = nothing
b_stat = nothing
buf = nothing
emptyf = nothing
f = nothing
s = nothing
gc()
gc()

but could not get consistent success here. Any suggestions about what might still be open and how to absolutely close it before calling ispath on the deleted file and directory? Skipping the test is a workaround, not a real solution.

[vtjnash]

@JeffBezanson libuv uses a number of flags to CreateFile to mitigate this. would you mind copying over their work to ios?

[JeffBezanson]

No problem; can you point me in the right direction?

[vtjnash]

somewhere in libuv/src/win/file.c. I haven't looked in a few months for the exact location

[tkelman]

Recently looked into the temporary folder that's getting used here, and found these tests are leaving around a lot of julia# temp files. Deletion of these temp files apparently isn't happening properly.

[vtjnash]

no, we don't make any effort to delete those, and neither does windows

[tkelman]

Huh? Isn't rm(file) an effort to delete? Looking into it the temp files are actually coming from the spawn tests, not file, so I misspoke there.

[vtjnash]

Oh, I guess we do make some effort then. It sounds like we are just inconsistent then.

[Keno]

@quinnj You had started to look into this. Any progress? Otherwise I'll take a look.

[quinnj]

No, I wasn't sure where to find libuv's CreateFile flags.

[Keno]

Ok.

[Keno]

I did the libuv thing, but it didn't actually help. I'm also not sure we have any open references around (which was the original issue). Interestingly this also doesn't fail for me at the REPL.

[Keno]

Ok, I straced this and it looks like that stat is failing with DELETE_PENDING, which is hardly surprising.

[Keno]

Anyway, turns out you can't detect DELETE_PENDING from user mode, so unless somebody wants to write a julia_delete kernel module, I say we just merge this pull request and live with it.

[tkelman]

Bummer, thanks for looking into it. Hopefully this won't be a pattern that comes up too frequently in packages or other user code that would lead to Windows-only breakage.

[tkelman]

Rebased.

[vtjnash]

@Keno ok to merge? Lgtm

[tkelman]

Thanks guys, another Windows test failure down. Underlying problem's still there, but apparently nothing we can do about it and I doubt it's all that consequential.