1 2007-04-23 Cluster File Systems, Inc. <info@clusterfs.com>
2 * version 1.4.11 / 1.6.1
3 * Support for networks:
4 socklnd - kernels up to 2.6.16
5 qswlnd - Qsnet kernel modules 5.20 and later
6 openiblnd - IbGold 1.8.2
8 viblnd - Voltaire ibhost 3.4.5 and later
10 iiblnd - Infiniserv 3.3 + PathBits patch
11 gmlnd - GM 2.1.22 and later
12 mxlnd - MX 1.2.1 or later
13 ptllnd - Portals 3.3 / UNICOS/lc 1.5.x, 2.0.x
16 ------------------------------------------------------------------------------
18 2007-04-01 Cluster File Systems, Inc. <info@clusterfs.com>
19 * version 1.4.10 / 1.6.0
20 * Support for networks:
21 socklnd - kernels up to 2.6.16
22 qswlnd - Qsnet kernel modules 5.20 and later
23 openiblnd - IbGold 1.8.2
25 viblnd - Voltaire ibhost 3.4.5 and later
26 ciblnd - Topspin 3.2.0
27 iiblnd - Infiniserv 3.3 + PathBits patch
28 gmlnd - GM 2.1.22 and later
29 mxlnd - MX 1.2.1 or later
30 ptllnd - Portals 3.3 / UNICOS/lc 1.5.x, 2.0.x
35 Description: gmlnd ignored some transmit errors when finalizing lnet messages.
39 Description: ptllnd logs a piece of incorrect debug info in kptllnd_peer_handle_hello.
43 Description: the_lnet.ln_finalizing was not set when the current thread is
44 about to complete messages. It only affects multi-threaded
50 Description: Changed the default kqswlnd ntxmsg=512
55 Description: Assertion failure in kernel ptllnd caused by posting passive
56 bulk buffers before connection establishment complete.
61 Description: A race in kernel ptllnd between deleting a peer and posting
62 new communications for it could hang communications -
63 manifesting as "Unexpectedly long timeout" messages.
68 Description: Kernel ptllnd lock ordering issue could hang a node.
73 Description: node crash on socket teardown race
76 Frequency : 'lctl peer_list' issued on a mx net
78 Description: Enable lctl's peer_list for MXLND
81 Frequency : after Ptllnd timeouts and portals congestion
83 Description: Credit overflows
84 Details : This was a bug in ptllnd connection establishment. The fix
85 implements better peer stamps to disambiguate connection
86 establishment and ensure both peers enter the credit flow
87 state machine consistently.
92 Description: kptllnd didn't propagate some network errors up to LNET
93 Details : This bug was spotted while investigating 11394. The fix
94 ensures network errors on sends and bulk transfers are
95 propagated to LNET/lustre correctly.
97 Severity : enhancement
99 Description: Fixed console chatter in case of -ETIMEDOUT.
101 Severity : enhancement
103 Description: Added D_NETTRACE for recording network packet history
104 (initially only for ptllnd). Also a separate userspace
105 ptllnd facility to gather history which should really be
106 covered by D_NETTRACE too, if only CDEBUG recorded history in
112 Description: o2iblnd handle early RDMA_CM_EVENT_DISCONNECTED.
113 Details : If the fabric is lossy, an RDMA_CM_EVENT_DISCONNECTED
114 callback can occur before a connection has actually been
115 established. This caused an assertion failure previously.
117 Severity : enhancement
119 Description: Multiple instances for o2iblnd
120 Details : Allow multiple instances of o2iblnd to enable networking over
121 multiple HCAs and routing between them.
125 Description: lnet deadlock in router_checker
126 Details : turned ksnd_connd_lock, ksnd_reaper_lock, and ksock_net_t:ksnd_lock
127 into BH locks to eliminate potential deadlock caused by
128 ksocknal_data_ready() preempting code holding these locks.
132 Description: Millions of failed socklnd connection attempts cause a very slow FS
133 Details : added a new route flag ksnr_scheduled to distinguish from
134 ksnr_connecting, so that a peer connection request is only turned
135 down for race concerns when an active connection to the same peer
136 is under progress (instead of just being scheduled).
138 ------------------------------------------------------------------------------
140 2007-02-09 Cluster File Systems, Inc. <info@clusterfs.com>
142 * Support for networks:
143 socklnd - kernels up to 2.6.16
144 qswlnd - Qsnet kernel modules 5.20 and later
145 openiblnd - IbGold 1.8.2
147 viblnd - Voltaire ibhost 3.4.5 and later
148 ciblnd - Topspin 3.2.0
149 iiblnd - Infiniserv 3.3 + PathBits patch
150 gmlnd - GM 2.1.22 and later
151 mxlnd - MX 1.2.1 or later
152 ptllnd - Portals 3.3 / UNICOS/lc 1.5.x, 2.0.x
155 Severity : major on XT3
157 Description: libcfs overwrites /proc/sys/portals
158 Details : libcfs created a symlink from /proc/sys/portals to
159 /proc/sys/lnet for backwards compatibility. This is no
160 longer required and makes the Cray portals /proc variables
165 Description: OFED FMR API change
166 Details : This changes parameter usage to reflect a change in
167 ib_fmr_pool_map_phys() between OFED 1.0 and OFED 1.1. Note
168 that FMR support is only used in experimental versions of the
169 o2iblnd - this change does not affect standard usage at all.
171 Severity : enhancement
173 Description: new ko2iblnd module parameter: ib_mtu
174 Details : the default IB MTU of 2048 performs badly on 23108 Tavor
175 HCAs. You can avoid this problem by setting the MTU to 1024
176 using this module parameter.
178 Severity : enhancement
179 Bugzilla : 11118/11620
180 Description: ptllnd small request message buffer alignment fix
181 Details : Set the PTL_MD_LOCAL_ALIGN8 option on small message receives.
182 Round up small message size on sends in case this option
183 is not supported. 11620 was a defect in the initial
184 implementation which effectively asserted all peers had to be
185 running the correct protocol version which was fixed by always
186 NAK-ing such requests and handling any misalignments they
191 Description: When kib(nal|lnd)_del_peer() is called upon a peer whose
192 ibp_tx_queue is not empty, kib(nal|lnd)_destroy_peer()'s
193 'LASSERT(list_empty(&peer->ibp_tx_queue))' will fail.
195 Severity : enhancement
197 Description: Patchless ZC(zero copy) socklnd
198 Details : New protocol for socklnd, socklnd can support zero copy without
199 kernel patch, it's compatible with old socklnd. Checksum is
200 moved from tunables to modparams.
204 Description: When ksocknal_del_peer() is called upon a peer whose
205 ksnp_tx_queue is not empty, ksocknal_destroy_peer()'s
206 'LASSERT(list_empty(&peer->ksnp_tx_queue))' will fail.
209 Frequency : when ptlrpc is under heavy use and runs out of request buffer
211 Description: In lnet_match_blocked_msg(), md can be used without holding a
215 Frequency : very rarely
217 Description: If ksocknal_lib_setup_sock() fails, a ref on peer is lost.
218 If connd connects a route which has been closed by
219 ksocknal_shutdown(), ksocknal_create_routes() may create new
220 routes which hold references on the peer, causing shutdown
221 process to wait for peer to disappear forever.
223 Severity : enhancement
225 Description: Dump XT3 portals traces on kptllnd timeout
226 Details : Set the kptllnd module parameter "ptltrace_on_timeout=1" to
227 dump Cray portals debug traces to a file. The kptllnd module
228 parameter "ptltrace_basename", default "/tmp/lnet-ptltrace",
229 is the basename of the dump file.
232 Frequency : infrequent
234 Description: kernel ptllnd fix bug in connection re-establishment
235 Details : Kernel ptllnd could produce protocol errors e.g. illegal
236 matchbits and/or violate the credit flow protocol when trying
237 to re-establish a connection with a peer after an error or
240 Severity : enhancement
242 Description: Allow /proc/sys/lnet/debug to be set symbolically
243 Details : Allow debug and subsystem debug values to be read/set by name
244 in addition to numerically, for ease of use.
247 Frequency : only in configurations with LNET routers
249 Description: routes automatically marked down and recovered
250 Details : In configurations with LNET routers if a router fails routers
251 now actively try to recover routes that are down, unless they
252 are marked down by an administrator.
254 ------------------------------------------------------------------------------
256 2006-12-09 Cluster File Systems, Inc. <info@clusterfs.com>
259 Frequency : very rarely, in configurations with LNET routers and TCP
261 Description: incorrect data written to files on OSTs
262 Details : In certain high-load conditions incorrect data may be written
263 to files on the OST when using TCP networks.
265 ------------------------------------------------------------------------------
267 2006-07-31 Cluster File Systems, Inc. <info@clusterfs.com>
269 - rework CDEBUG messages rate-limiting mechanism b=10375
270 - add per-socket tunables for socklnd if the kernel is patched b=10327
272 ------------------------------------------------------------------------------
274 2006-02-15 Cluster File Systems, Inc. <info@clusterfs.com>
276 - fix use of portals/lnet pid to avoid dropping RPCs b=10074
277 - iiblnd wasn't mapping all memory, resulting in comms errors b=9776
278 - quiet LNET startup LNI message for liblustre b=10128
279 - Better console error messages if 'ip2nets' can't match an IP address
280 - Fixed overflow/use-before-set bugs in linux-time.h
281 - Fixed ptllnd bug that wasn't initialising rx descriptors completely
282 - LNET teardown failed an assertion about the route table being empty
283 - Fixed a crash in LNetEQPoll(<invalid handle>)
284 - Future protocol compatibility work (b_rls146_lnetprotovrsn)
285 - improve debug message for liblustre/Catamount nodes (b=10116)
287 2005-10-10 Cluster File Systems, Inc. <info@clusterfs.com>
288 * Configuration change for the XT3
289 The PTLLND is now used to run Lustre over Portals on the XT3.
290 The configure option(s) --with-cray-portals are no longer
291 used. Rather --with-portals=<path-to-portals-includes> is
292 used to enable building on the XT3. In addition to enable
293 XT3 specific features the option --enable-cray-xt3 must be
296 2005-10-10 Cluster File Systems, Inc. <info@clusterfs.com>
297 * Portals has been removed, replaced by LNET.
298 LNET is new networking infrastructure for Lustre, it includes a
299 reorganized network configuration mode (see the user
300 documentation for full details) as well as support for routing
301 between different network fabrics. Lustre Networking Devices
302 (LNDS) for the supported network fabrics have also been created
303 for this new infrastructure.
305 2005-08-08 Cluster File Systems, Inc. <info@clusterfs.com>
310 Frequency : rare (large Voltaire clusters only)
312 Description: the default number of reserved transmit descriptors was too low
313 for some large clusters
314 Details : As a workaround, the number was increased. A proper fix includes
317 2005-06-02 Cluster File Systems, Inc. <info@clusterfs.com>
322 Frequency : occasional (large-scale events, cluster reboot, network failure)
324 Description: too many error messages on console obscure actual problem and
325 can slow down/panic server, or cause recovery to fail repeatedly
326 Details : enable rate-limiting of console error messages, and some messages
327 that were console errors now only go to the kernel log
329 Severity : enhancement
331 Description: add /proc/sys/portals/catastrophe entry which will report if
332 that node has previously LBUGged
334 2005-04-06 Cluster File Systems, Inc. <info@clusterfs.com>
336 - update gmnal to use PTL_MTU, fix module refcounting (b=5786)
338 2005-04-04 Cluster File Systems, Inc. <info@clusterfs.com>
340 - handle error return code in kranal_check_fma_rx() (5915,6054)
342 2005-02-04 Cluster File Systems, Inc. <info@clusterfs.com>
344 - update vibnal (Voltaire IB NAL)
345 - update gmnal (Myrinet NAL), gmnalid
347 2005-02-04 Eric Barton <eeb@bartonsoftware.com>
349 * Landed portals:b_port_step as follows...
351 - removed CFS_DECL_SPIN*
352 just use 'spinlock_t' and initialise with spin_lock_init()
354 - removed CFS_DECL_MUTEX*
355 just use 'struct semaphore' and initialise with init_mutex()
357 - removed CFS_DECL_RWSEM*
358 just use 'struct rw_semaphore' and initialise with init_rwsem()
360 - renamed cfs_sleep_chan -> cfs_waitq
361 cfs_sleep_link -> cfs_waitlink
363 - fixed race in linux version of arch-independent socknal
364 (the ENOMEM/EAGAIN decision).
366 - Didn't fix problems in Darwin version of arch-independent socknal
367 (resetting socket callbacks, eager ack hack, ENOMEM/EAGAIN decision)
369 - removed libcfs types from non-socknal header files (only some types
370 in the header files had been changed; the .c files hadn't been