-----BEGIN PGP SIGNATURE-----
iHUEABYKAB0WIQRAhzRXHqcMeLMyaSiRxhvAZXjcogUCY5bw/wAKCRCRxhvAZXjc
ol79AQCsHS9s78dLUvdasfQ1023dyF9zaQ8XGkDO6tRssJzGAAD7B8odxDsfQgjQ
Qzzn9YPZVUgHjd4xBg21UVPmRP5snwQ=
=wYgr
-----END PGP SIGNATURE-----
Merge tag 'fs.xattr.simple.rework.rbtree.rwlock.v6.2' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/idmapping
Pull simple-xattr updates from Christian Brauner:
"This ports the simple xattr infrastucture to rely on a simple rbtree
protected by a read-write lock instead of a linked list protected by a
spinlock.
A while ago we received reports about scaling issues for filesystems
using the simple xattr infrastructure that also support setting a
larger number of xattrs. Specifically, cgroups and tmpfs.
Both cgroupfs and tmpfs can be mounted by unprivileged users in
unprivileged containers and root in an unprivileged container can set
an unrestricted number of security.* xattrs and privileged users can
also set unlimited trusted.* xattrs. A few more words on further that
below. Other xattrs such as user.* are restricted for kernfs-based
instances to a fairly limited number.
As there are apparently users that have a fairly large number of
xattrs we should scale a bit better. Using a simple linked list
protected by a spinlock used for set, get, and list operations doesn't
scale well if users use a lot of xattrs even if it's not a crazy
number.
Let's switch to a simple rbtree protected by a rwlock. It scales way
better and gets rid of the perf issues some people reported. We
originally had fancier solutions even using an rcu+seqlock protected
rbtree but we had concerns about being to clever and also that
deletion from an rbtree with rcu+seqlock isn't entirely safe.
The rbtree plus rwlock is perfectly fine. By far the most common
operation is getting an xattr. While setting an xattr is not and
should be comparatively rare. And listxattr() often only happens when
copying xattrs between files or together with the contents to a new
file.
Holding a lock across listxattr() is unproblematic because it doesn't
list the values of xattrs. It can only be used to list the names of
all xattrs set on a file. And the number of xattr names that can be
listed with listxattr() is limited to XATTR_LIST_MAX aka 65536 bytes.
If a larger buffer is passed then vfs_listxattr() caps it to
XATTR_LIST_MAX and if more xattr names are found it will return
-E2BIG. In short, the maximum amount of memory that can be retrieved
via listxattr() is limited and thus listxattr() bounded.
Of course, the API is broken as documented on xattr(7) already. While
I have no idea how the xattr api ended up in this state we should
probably try to come up with something here at some point. An iterator
pattern similar to readdir() as an alternative to listxattr() or
something else.
Right now it is extremly strange that users can set millions of xattrs
but then can't use listxattr() to know which xattrs are actually set.
And it's really trivial to do:
for i in {1..1000000}; do setfattr -n security.$i -v $i ./file1; done
And around 5000 xattrs it's impossible to use listxattr() to figure
out which xattrs are actually set. So I have suggested that we try to
limit the number of xattrs for simple xattrs at least. But that's a
future patch and I don't consider it very urgent.
A bonus of this port to rbtree+rwlock is that we shrink the memory
consumption for users of the simple xattr infrastructure.
This also adds kernel documentation to all the functions"
* tag 'fs.xattr.simple.rework.rbtree.rwlock.v6.2' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/idmapping:
xattr: use rbtree for simple_xattrs
02bf43c7b7