Hi,
when strace'ing "ls -l" on Linux, I see that it runs three syscalls for
every file in the directory, e.g.:
$ strace ls -l
...
lstat("test", {st_mode=S_IFREG|0644, st_size=0, ...}) = 0
lgetxattr("test", "security.selinux", 0x62d130, 255) = -1 ENODATA (No
data available)
getxattr("test", "system.posix_acl_access", 0x0, 0) = -1 EOPNOTSUPP
(Operation not supported)
lstat("test2", {st_mode=S_IFREG|0644, st_size=0, ...}) = 0
lgetxattr("test2", "security.selinux", 0x62b350, 255) = -1 ENODATA (No
data available)
getxattr("test2", "system.posix_acl_access", 0x0, 0) = -1 EOPNOTSUPP
(Operation not supported)
lstat("test3", {st_mode=S_IFREG|0644, st_size=0, ...}) = 0
lgetxattr("test3", "security.selinux", 0x62d150, 255) = -1 ENODATA (No
data available)
getxattr("test3", "system.posix_acl_access", 0x0, 0) = -1 EOPNOTSUPP
(Operation not supported)
...
The first operation is lstat(), which definitely makes sense. The other
two operations are
1) lgetxattr(..., "security.selinux", ...): Doesn't make sense imho,
because selinux isn't even enabled on the system on which I ran "ls -l".
2) getxattr(..., "system.posix_acl_access", ...): Doesn't make sense, as
ACLs are not supported on this file system (=> EOPNOTSUPP).
While these extra calls might not really be a problem for local file
systems, they are a real pain for remote file systems, such as NFS or
FhGFS (on which I'm working) and probably also for FUSE-based file systems.
The problem is that these extra syscalls lead to an inode revalidation
inside the (Linux) kernel and thus result in extra remote calls for
every file, so it would be really great if they could be avoided after
detecting that they don't make sense in a particular directory.
I don't want to start a philosophical discussion about large
directories, but as a matter of fact, we often have users with tens of
thousands of files inside a single directory and those users need to run
"ls -l" in these directories frequently to monitor the status of thei
compute job and things like that.
Currently, they need to wait three times longer than necessary to see
the result, due to the two extra syscalls for each file.
What I would like to suggest is detecting whether these extra calls make
sense and disable them if they don't make sense:
1) For lgetxattr(..., "security.selinux", ...), ls could simply check
whether selinux is enabled and skip the calls if it isn't enabled.
2) For getxattr(..., "system.posix_acl_access", ...), ls could check
whether the first call returns EOPNOTSUPP and then skip this call for
the remaining directory entries, because those will just have the same
return value.
Best regards,
Sven
---
Sven Breuner
Fraunhofer HPC
FhGFS Team
http://www.fhgfs.com
