2 ======================================================================
4 General ideas for improvements
5 ------------------------------
7 * Listen on specific interfaces or protocols (eg. only IPv6).
9 * Performance - measure and improve it.
11 * Exit on last connection (the default behaviour of qemu-nbd unless
14 * Limit number of incoming connections (like qemu-nbd -e).
16 * For parallel plugins, only create threads on demand from parallel
17 client requests, rather than pre-creating all threads at connection
18 time, up to the thread pool size limit. Of course, once created, a
19 thread is reused as possible until the connection closes.
21 * Async callbacks. The current parallel support requires one thread
22 per pending message; a solution with fewer threads would split
23 low-level code between request and response, where the callback has
24 to inform nbdkit when the response is ready:
25 https://www.redhat.com/archives/libguestfs/2018-January/msg00149.html
27 * More NBD protocol features. The currently missing features are
28 structured replies for sparse reads, block size constraints, and
31 * Add a callback to let plugins request minimum alignment for the
32 buffer to pread/pwrite; useful for a plugin utilizing O_DIRECT or
33 other situation where pre-aligned buffers are more efficient.
34 Ideally, a blocksize filter would honor strict alignment below and
35 advertise loose alignment above; all other filters (particularly
36 ones like offset) can fail to initialize if they can't guarantee
37 strict alignment and don't want to deal with bounce buffers.
39 * Test that zero-length read/write/extents requests behave sanely
40 (NBD protocol says they are unspecified).
42 * If a client negotiates structured replies, and issues a read/extents
43 call that exceeds EOF (qemu 3.1 is one such client, when nbdkit
44 serves non-sector-aligned images), return the valid answer for the
45 subset of the request in range and then NBD_REPLY_TYPE_ERROR_OFFSET
46 for the tail, rather than erroring the entire request.
48 * Test and document how to run nbdkit from inetd and xinetd in
51 * Audit the code base to get rid of strerror() usage (the function is
52 not thread-safe); however, using geterror_r() can be tricky as it
53 has a different signature in glibc than in POSIX.
55 * Teach nbdkit_error() to have smart newline appending (for existing
56 inconsistent clients), while fixing internal uses to omit the
57 newline. Commit ef4f72ef has some ideas on smart newlines, but that
58 should probably be factored into a utility function.
60 * We may need a way to lock nbdkit into memory and adjust the OOM
61 killer score. See this LKML discussion:
62 https://www.mail-archive.com/linux-kernel@vger.kernel.org/msg1933394.html
63 and also look at the implementation of the -swap option in
66 * Add a mode of operation where nbdkit is handed a pre-opened fd to be
67 used immediately in transmission phase (skipping handshake). There
68 are already third-party clients of the kernel's /dev/nbdX which rely
69 on their own protocol instead of NBD handshake, before calling
70 ioctl(NBD_SET_SOCK); this mode would let the third-party client
71 continue to keep their non-standard handshake while utilizing nbdkit
72 to prototype new behaviors in serving the kernel.
74 * Clients should be able to list export names supported by plugins.
75 Current behaviour is not really correct: We only list the -e
76 parameter from the command line, which is different from the export
77 name(s) that a plugin might want to support. Probably we should
78 deprecate the -e option entirely since it does nothing useful.
80 * Add plugin "connect" method. This would be called on a connection
81 before handshaking or TLS negotiation, and could be used (with
82 nbdkit_peer_name) to accept or reject connections based on IP
83 address, rather like a poor man's TCP wrappers. See also commit
86 * Background thread for filters. Some filters (readahead, cache and
87 proposed scan filter - see below) could be more effective if they
88 were able to defer work to a background thread. However it's not as
89 simple as just creating a background thread, because you only have
90 access to the connection / next_ops from the context of a filter
91 callback. (Also you can't "save" next_ops because it becomes
92 invalid outside the callback). The background thread would need to
93 have its own connection to the plugin, which would be independent of
94 any client connection, and this requires some care because it breaks
95 an existing assumption.
97 * Add scan filter. This would be placed on top of cache filters and
98 would scan (read) the whole disk in the background, ensuring it is
99 copied into the cache. Useful if you have a slow plugin, limited
100 size device, and lots of local disk space, especially if you know
101 that the NBD clients will eventually read all of the device. RWMJ
102 wrote an implementation of this but it doesn't work well without a
105 Suggestions for plugins
106 -----------------------
108 Note: qemu supports other formats such as libnfs, iscsi, gluster and
109 ceph/rbd, and while similar plugins could be written for nbdkit there
110 is no compelling reason unless the result is better than qemu-nbd.
111 For the majority of users it would be better if they were directed to
112 qemu-nbd for these use cases.
116 https://lists.gnu.org/archive/html/qemu-devel/2017-11/msg02971.html
117 is a partial solution but it needs cleaning up.
119 nbdkit-floppy-plugin:
121 * Add boot sector support. In theory this is easy (eg. using
122 SYSLINUX), but the practical reality of making a fully bootable
123 floppy is rather more complex.
125 * Add multiple dir merging.
127 nbdkit-linuxdisk-plugin:
129 * Add multiple dir merging (in e2fsprogs mke2fs).
131 Suggestions for filters
132 -----------------------
134 * tar plugin should really be a filter
136 * gzip plugin should really be a filter
138 * libarchive could be used to implement a general tar/zip filter
140 * LUKS encrypt/decrypt filter, bonus points if compatible with qemu
141 LUKS-encrypted disk images
143 * masking plugin features for testing clients (see 'nozero' and 'fua'
144 filters for examples)
146 * nbdkit-cache-filter should handle ENOSPC errors automatically by
147 reclaiming blocks from the cache
149 * zstd filter was requested as a way to do what we currently do with
150 xz but saving many hours on compression (at the cost of hundreds of
152 https://github.com/facebook/zstd/issues/395#issuecomment-535875379
156 * allow other kinds of traffic shaping such as VBR
158 * limit traffic per client (ie. per IP address)
160 * split large requests to avoid long, lumpy sleeps when request size
161 is much larger than rate limit
165 * allow user to specify which errors cause a retry and which ones are
166 passed through; for example there's probably no point retrying on
169 * implement a softer mode (retry-reopen=no) where we don't reopen the
170 plugin, we just retry the data command that failed
172 * there are all kinds of extra complications possible here,
173 eg. specifying a pattern of retrying and reopening:
174 retry-method=RRRORRRRRORRRRR meaning to retry the data command 3
175 times, reopen, retry 5 times, etc.
182 Things like blacklisting or whitelisting IP addresses can be done
183 using external wrappers (TCP wrappers, systemd).
185 However it might be nice to have a configurable filter for preventing
186 valid but not sensible requests. The server already filters invalid
187 requests. This would be like seccomp, and could be implemented using
188 an eBPF-based parser. Unfortunately actual eBPF is difficult to use
189 for userspace processes. The "standard" isn't solidly defined - the
190 Linux kernel implementation is the standard - and Linux has by far the
191 best implementation, particularly around bytecode verification and
192 JITting. There is a userspace VM (ubpf) but it has very limited
193 capabilities compared to Linux.
198 Filters allow certain types of composition, but others would not be
199 possible, for example RAIDing over multiple nbd sources. Because the
200 plugin API limits us to loading a single plugin to the server, the
201 best way to do this (and the most robust) is to compose multiple
202 nbdkit processes. Perhaps libnbd will prove useful for this purpose.
207 * Figure out how to get 'make distcheck' working. VPATH builds are
208 working, but various pkg-config results that try to stick
209 bash-completion and ocaml add-ons into their system-wide home do
210 not play nicely with --prefix builds for a non-root user.
217 * Consider supporting a more idiomatic style for writing Rust plugins.
219 * Better documentation.
223 * There is no attempt to ‘make install’ or otherwise package the
224 crate. Since it looks as if Rust code is normally distributed as
225 source it's not clear what that would even mean.
230 From time to time we may update the plugin protocol. This section
231 collects ideas for things which might be fixed in the next version of
234 Note that we keep the old protocol(s) around so that source
235 compatibility is retained. Plugins must opt in to the new protocol
236 using ‘#define NBDKIT_API_VERSION <version>’.
238 * All methods taking a ‘count’ field should be uint64_t (instead of
239 uint32_t). Although the NBD protocol does not support 64 bit
240 lengths, it might do in future.
242 * pread could be changed to allow it to support Structured Replies
243 (SRs). This could mean allowing it to return partial data, holes,
244 zeroes, etc. For a client that negotiates SR coupled with a plugin
245 that supports .extents, the v2 protocol would allow us to at least
246 synthesize NBD_REPLY_TYPE_OFFSET_HOLE for less network traffic, even
247 though the plugin will still have to fully populate the .pread
248 buffer; the v3 protocol should make sparse reads more direct.
250 * Parameters should be systematized so that they aren't just (key,
251 value) strings. nbdkit should know the possible keys for the plugin
252 and filters, and the type of the values, and both check and parse
255 * Modify open() API so it takes an export name parameter.