Pádraig Brady <[email protected]> writes: > On 14/11/2026 07:38, Collin Funk wrote: >> Thai Duong <[email protected]> writes: >> >>> Thanks, will do! >> And thanks to you for the report! It was reported by Anthropic's OSS >> scanner as well, which is interesting. I'll reference both reports in >> the commit message. >> I've attached a patch that should fix the issue. However, I'm going >> to >> hold off on pushing until I write some additional tests tomorrow, since >> the changes mean chown/chmod can be called from two separate places, one >> of which is shared with 'cp' and 'mv'. >> Avoiding this race, and some other ones around this area, means >> using >> reflinks less on macOS, which I am not super happy about. As far as I >> know, Linux is the only operating system with the ability to reflink >> from two file descriptors. The macOS API is marginally better than >> Solaris since you can reflink from a source file descriptor, but you >> still need to open the destination afterward, which may be replaced with >> something else. > This change looks good. > > There is a new dependency where the lchmodat() used with `install -s` > introduces a new dependency on /proc being available on glibc < 2.39, > or on kernels < 6.6, but I have a separate patch to gnulib to provide > a fallback on such systems, so I think this patch is fine. > > This patch also changes things so the final permission is set (say +x) > in the case where close() fails due to a delayed EIO. > But I don't think that's a practical concern as install(1) doesn't > provide atomic replacement of the destination anyway, > which would be the way to robustly handle this and other races.
Thanks for the review. I pushed it now, after noticing there's some more --strip than I remember, which should catch the bugs with setting modes and such. We can have a look at adding some tests for the races with FIFOs or 'strace' separately. Marking this one as done. Collin
