An issue was reported today on the Alpine Linux tracker at
https://gitlab.alpinelinux.org/alpine/aports/issues/10960 regarding
readdir results from SMB/NFS shares with musl libc.
After a good deal of analysis, we determined the root cause to be that
the second and subsequent calls to getdents64 are dropping/skipping
direntries (that have not yet been deleted) when some entries were
deleted following the previous call. The issue appears to happen only
when the buffer size passed to getdents64 is below some threshold
greater than 2k (the size musl uses) but less than 32k (the size glibc
uses, with which we were unable to reproduce the issue).
My guess at the mechanism of failure is that the kernel has cached
some entries which it obtained from the FS server based on whatever
its preferred transfer size is, but didn't yet pass them to userspace
due to limited buffer space, and then purged the buffer when resuming
getdents64 after some entries were deleted for reasons related to the
changes made way back in 0c0308066ca5 (NFS: Fix spurious readdir
cookie loop messages). If so, any such purge likely needs to be
delayed until already-buffered results are read, and there may be
related buggy interactions with seeking that need to be examined.
The to/cc for this message are just my best guesses. Please cc anyone
I missed who should be included when replying, and keep me on cc since
I'm not subscribed to any of these lists but the musl one.
Rich
* Rich Felker:
> An issue was reported today on the Alpine Linux tracker at
> https://gitlab.alpinelinux.org/alpine/aports/issues/10960 regarding
> readdir results from SMB/NFS shares with musl libc.
>
> After a good deal of analysis, we determined the root cause to be that
> the second and subsequent calls to getdents64 are dropping/skipping
> direntries (that have not yet been deleted) when some entries were
> deleted following the previous call. The issue appears to happen only
> when the buffer size passed to getdents64 is below some threshold
> greater than 2k (the size musl uses) but less than 32k (the size glibc
> uses, with which we were unable to reproduce the issue).
From the Gitlab issue:
while ((dp = readdir(dir)) != NULL) {
unlink(dp->d_name);
++file_cnt;
}
I'm not sure that this is valid code to delete the contents of a
directory. It's true that POSIX says this:
| If a file is removed from or added to the directory after the most
| recent call to opendir() or rewinddir(), whether a subsequent call
| to readdir() returns an entry for that file is unspecified.
But many file systems simply provide not the necessary on-disk data
structures which are need to ensure stable iteration in the face of
modification of the directory. There are hacks, of course, such as
compacting the on-disk directory only on file creation, which solves
the file removal case.
For deleting an entire directory, that is not really a problem because
you can stick another loop around this while loop which re-reads the
directory after rewinddir. Eventually, it will become empty.