add a suffix to a new cache files in case of failure of renaming it to an exisiting cache file already in use

[KristofferC]

In cases where we try to compile a cache file for a package that is already in use in another Julia session, Windows gives us an error because we are trying to rename a file to another file that is already in use (#48074).

This PR detects when the rename fails and adds a suffix _i to the cache file name where i is the smallest integer such that that file does not exist.

Starting a few julia sessions and loading Example while editing it between every load gives a list of files like:

PS C:\Users\Kristoffer\.julia\compiled\v1.10\Example> dir

    Directory: C:\Users\Kristoffer\.julia\compiled\v1.10\Example
Mode                 LastWriteTime         Length Name
----                 -------------         ------ ----
-a---          2023-01-05    14:11          20480 lLvWP_7lexV_1.dll
-a---          2023-01-05    14:11           2732 lLvWP_7lexV_1.ji
-a---          2023-01-05    14:12          20480 lLvWP_7lexV_2.dll
-a---          2023-01-05    14:12           2738 lLvWP_7lexV_2.ji
-a---          2023-01-05    14:11          20480 lLvWP_7lexV.dll
-a---          2023-01-05    14:11           2728 lLvWP_7lexV.ji

Fixes #48074

[gbaraldi]

Why are we trying to overwrite the cache files? If they aren't different shouldn't they have different hashes?

[KristofferC]

There is a maximum of 10 cache files for a package before we start overwriting old files. If you are continuously working on a package that means you would quickly start deleting cache files used by other projects. In the example above, I would have deleted the cache file for Example that I use in my global project. So that is why a cache file for a given project will overwrite an earlier cachefile from the same project.

But there might be better strategies and heuristics.

[vchuravy]

LGTM, but I would love to have the POSIX semantics that C++ now mandates in libuv ;)

[vtjnash]

Suggested change

	old_cachefiles = filter(x->startswith(x, basename(ocachename)) && endswith(x, ocacheext), readdir(cachepath))
	nums = Set(1:length(old_cachefiles))
	for file in old_cachefiles
	name = splitext(file)[1]
	s = split(name, '_')
	if length(s) == 3 # e.g. lLvWP_g5TNZ_3
	i = tryparse(Int, last(s))
	if i !== nothing
	delete!(nums, i)
	end
	end
	end
	# TODO: Risk for a race here if some other process grabs this name before us
	num = isempty(nums) ? 1 : minimum(nums)
	ocachefile = ocachename * "_$num" * ocacheext
	old_cachefiles = Set(readdir(cachepath))
	num = 1
	while true
	ocachefile = ocachename * "_$num" * ocacheext
	# ispath(ocachefile) || break
	in(ocachefile, old_cachefiles) || break
	end
	# TODO: Risk for a race here if some other process grabs this name before us

[KristofferC]

Yeah, this part felt a bit convoluted.

[vtjnash]

We could also try to use tempname, since it implements mostly the same logic as this, but possibly does not support suffixes

[DilumAluthge]

Backport to 1.9?

[grandinj]

Note that Win32 and Linux have different semantics and API around atomic filesystem operations.

Ideally you want to be using the ReplaceFile function on windows.

[vtjnash]

You don't though, since that function is pretty weird and not like posix nor atomic. You want MoveFile, which is what this calls.