worker: avoid potential partial write of pickled data
Previously, the code wrote the pickled data using os.write(). However,
os.write() can write less bytes than passed to it. To trigger the problem, the
pickled data had to be larger than
2147479552 bytes on my system.
Instead, open a file object and pass it to pickle.dump(). This also has the
advantage that it doesn’t buffer the whole pickled data in memory.
Note that the opened file must be buffered because pickle doesn’t support
unbuffered streams because unbuffered streams’ write() method might write less
bytes than passed to it (like os.write()) but pickle.dump() relies on that all
bytes are written (see https://github.com/python/cpython/issues/93050).
The side effect of using a file object and a with statement is that wfd is
explicitly closed now while it seems like before it was implicitly closed by
process exit.
#require no-windows
$ . "$TESTDIR/remotefilelog-library.sh"
$ hg init master
$ cd master
$ cat >> .hg/hgrc <<EOF
> [remotefilelog]
> server=True
> serverexpiration=-1
> EOF
$ echo x > x
$ hg commit -qAm x
$ echo y > y
$ hg commit -qAm y
$ echo z > z
$ hg commit -qAm z
$ cd ..
$ hgcloneshallow ssh://user@dummy/master shallow -q
3 files fetched over 1 fetches - (3 misses, 0.00% hit ratio) over *s (glob)
# Compute keepset for 0th and 2nd commit, which implies that we do not process
# the 1st commit, therefore we diff 2nd manifest with the 0th manifest and
# populate the keepkeys from the diff
$ cd shallow
$ cat >> .hg/hgrc <<EOF
> [remotefilelog]
> pullprefetch=0+2
> EOF
$ hg debugkeepset
# Compute keepset for all commits, which implies that we only process deltas of
# manifests of commits 1 and 2 and therefore populate the keepkeys from deltas
$ cat >> .hg/hgrc <<EOF
> [remotefilelog]
> pullprefetch=all()
> EOF
$ hg debugkeepset