Quoting Benjamin Thery (benjamin.thery@xxxxxxxx): > Greg, > > Here is an updated version of the sysfs tagged directories that improves > a bit the situation over the previous one. > > I've modified the patch 09 ("Enable tagging for net_class directories in > sysfs") to be a bit less intrusive in sysfs core. I removed the #ifdef'd > parts you didn't like in fs/sysfs/mount.c, and replaced it by a generic > routine sysfs_ns_exit() that is called, if needed, by the namespace when > it exits. This routine goes through every sysfs super blocks and calls > the callback passed by the namespace to clean its tag. > > The patch is now splitted in two: > * 09/11: the generic routine, > * 10/11: the remaining network parts. > > The generic part can be merge with patch 05 ("sysfs: Implement sysfs > tagged directory support.") but I left it separate for now to ease > reviews. Thanks. Not nearly as radical as I'd feared. Greg, does this address your concern? It stops the need to put any non-braindead code (i.e. the _exit notifier stuff) in fs/sysfs/mount.c for any new namespaces needing sysfs tagging, leaving only simple comparison/assignments. > Serge's patch for user namespace is modified to use this new service too. > No more #ifdef CONFIG_NET or #ifdef CONFIG_USER_NS in fs/sysfs/mount.c > now. > > > But, currently, a new namespace that wants to add its tag to sysfs dirs > still need to modify fs/sysfs/mount.c in a few routines to manage the > new tag member added in struct sysfs_tag_info: sysfs_fill_super(), > sysfs_test_super(), sysfs_kill_sb() (see the last two patches). > These changes are only the initialization and a bunch of comparisons. > > If we really want to go further, to get rid of these, I've thought > about: > > * Extending sysfs_tagged_dir_operations with super blocks operations: > - fill_sb_tag, test_sb_tag, kill_sb_tag > > * Add routines in sysfs to allow registration/unregistration of these > operations structs in a list: > - sysfs_register_tagged_dir_ops()... > > * Each subsystem concerned implements and registers its operations at > boot. > > * In sysfs_fill_super(), sysfs_test_super() and sysfs_kill_sb(), add > loops to go through all registered operations structs and calls the > corresponding operations if it's present. > > But... I thought it was a bit overkill for the few namespaces that will > actually need sysfs tagged directories. Agreed. We expect device namespaces... anything else? thanks, -serge _______________________________________________ Containers mailing list Containers@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linux-foundation.org/mailman/listinfo/containers