#166 OpenSSI fails the glibc tst-atime test

[John Hughes]

CFS doesn't keep the atime (last accessed time) field up to date.

This is a known limitation, I'm just reporting it here so it doesn't get forgotten about.

Maybe as a workaround we could pretend that cfs filesystems were mounted with the noatime option?

To reproduce:

$ cc tst-atime.c
$ ./a.out
atime has not changed

[John Hughes]

[John Hughes]

Simple idea, just do this at user level by remounting the fs with an explicit noatime option:

mount -o remount,noatime /

doesn't work because statvfs reads /proc/mounts to find the mount options and some of the weird crud we have hanging around confuses it:

# cat /proc/mounts
rootfs / rootfs rw 0 0
/dev2/root2 / cfs rw,noatime,chard,node=1 0 0
...

Which is the real root? Apparently statvfs thinks it's the first one.

[John Hughes]

Actually the:
rootfs / rootfs
stuff is not specific to OpenSSI.

What happens in statvfs is it looks for a mountpoint with the same filesystem type as the file it's been given (so it's looking for ext2/ext3) but from /proc our fs shows up as cfs, so it doesn't see that it's the same one.

Later on it retries without the type, but this time around the "rootfs / rootfs" one matches and that doesn't have the flags.

[John Hughes]

What's very bizarre is that this sometimes works on nodes other than the one where the underlying filesystem is mounted, the access time seems to get changed the first time a file is read, but not afterwards.

[Roger Tsang]

Testing a patch. So far successfully tested at server.

[Nobody/Anonymous]

Tracker.. Huh, really? :)