diff options
Diffstat (limited to 'linux-nfs/TODO')
| -rw-r--r-- | linux-nfs/TODO | 121 |
1 files changed, 121 insertions, 0 deletions
diff --git a/linux-nfs/TODO b/linux-nfs/TODO new file mode 100644 index 0000000..3a439d4 --- /dev/null +++ b/linux-nfs/TODO @@ -0,0 +1,121 @@ + +Todo/Status List for Linux-NFS + + * denotes to be done; + o denotes draft implementation, possibly commented out + - denotes done, + + denotes done and tested +------------------------------------------------------------------ + +RPC: + + * Server-side AUTH_DES authentication + +NFS: + + * stat() calls don't check whether the cached attrs are stil valid + (this is a problem in the VFS). + - NFS_ROOT stuff needs fixing. + o Swapping over NFS. + + Issues of swapout: + * Avoid recursion in low memory situations where + kmalloc may call try_to_swap_out etc ad inf. + * Don't do async I/O on swap files. + + For special-casing related to NFS swap I/O, flag swap file + semantics in inode->i_flags. In swapfile.c, change functions + to call readpage/writepage if available, otherwise proceed + as usual. + + - Write-back support. + * Disable page cache invalidation/flushing for locked file + regions. + - Directory caching (we now have page-sized dircache entries + which could easily be organized into a linked list). These + dircache pages come along as a linked list that can be copied + almost 1-to-1 into a dirent struct. If this is put into the + VFS, other remote fs's will also benefit. + + [Note: I just increased the readdir cache to hold more than + one directory. With this, the exclusive lock on readdir goes + away, too. With a larger cache, it may also be worth to think + about directory readahead...] + * Better lookup caching? + * When a read lock is present, don't time out attr cache or + page cache for that region. Likewise, if a write lock is present, + be lazy on write-back. + * Implement CTO. + - BUG: Invalidate readdir cache after remove/rename/unlink + * Automatic `mounting' when the server crosses mount points + transparently (some IRIX machines seem to do this when + using -nohide). + * NFSv3 support. This requires careful design to maximize + code sharing between NFSv2 and NFSv3. + * More robust rename handling (see comment before nfs_rename). + * Add Miquel's O_EXCL hack for file creation. + * Performance improvement: When a complete reply is received, and + the (async) task is woken up, don't put it on rpciod's scheduling + queue, but add it to a `fast scheduler queue.' The fast scheduler + could be a special handler that's registered on the tq_scheduler task + queue. This queue is fired by the kernel scheduler as soon as + the other bottom halves have been run. + + Note that implementing this for sync tasks is even trickier than + for async tasks, because you have to make sure you do the right + thing in rpc_sleep_on(). + * writeback of writable mmaps. Dirty pages are not subject to + writeback scheduling. Also, msync should make sure pages are + written with O_SYNC on. + + +nfsd: + + * uid/gid mapping, and rpc.ugidd support + - Don't read/write a file that might have mandatory locks. + * Implement secure/kerberos export options (take care of lockd + fopen() calls--most clients seem to use NULL creds for lockd). + - there's a bug in readdir wrt large directories. Try mounting + the linux source tree and do an ls on include/linux... + * Support for UNIX socket creation. + * Someone should look over the error return codes. I tend to + mix up EPERM and EACCES. + * NFSv3 support. + - Refuse to look up inodes in procfs (security issues). + o Delayed writes (delay syncing of file data when nfsd handles + several write requests for the same file concurrently). + (Draft - see nfsd_write in fs/nfsd/write.c. Needs benchmarking). + * Faster read operations (single copy): mmap the file region + to be read into VM, and pass the VMA to the xdr routines + which pass the region's VM address into sock->ops->writemsg. + This copies the file data directly from the page cache into + the network buffer. + Release the vma region after encoding. + * Faster write operations (single copy, with IPv6 net layout): + Get the unfragmented UDP datagram, pull the header and + do normal processing. Then mmap the file, copy the write data, + and release VMA. + - Clear setuid/setgid bit after write(). + * Quota support. + +lockd: + + * Server should run on privileged port. + * Testing reclaim support. + * HP lockd accepts our GRANT_MSG callback and passes on the grant + to the blocking process, but doesn't reply with a GRANT_RES. + It's not clear to me why it would do this. + * Unregister hosts (SM_UNMON) with rpc.statd when appropriate. + +mountd + + * Unregister service from portmapper upon exit/SIGTERM + +mount + + * If available, use version 3 of the mount protocol and + obtain pathconf data (fill in data->bsize). + +documentation: + + - Manpages need to be written |