Versions Compared

Old Version 8

changes.mady.by.user Amir Shehata

Saved on

compared with

New Version 9

changes.mady.by.user Amir Shehata

Saved on

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

https://review.whamcloud.com/#/c/33447/

More Details
Code Block
LU-10153 lnet: remove route add restriction
    
Remove restriction with adding routes to the same remote network
via two different gateways.
    
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: Iefc5aa10f73e9e7bdd283f5e933fbb8ee819df50

There is no need to restrict the addition of routes to the same remote network via two different gateways.

This change is simple. Just remove lnet_check_routes() and its callers.

https://review.whamcloud.com/#/c/33182
Code Block
LU-11292 lnet: Discover routers on first use

Discover routers on first use. This brings the behavior when
interacting with routers inline with when dealing with normal
peers.

Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: I8527e41daf2f5f6ab5f04aac1285aaa6cc4ee594
  • lnet_initiate_peer_discovery()
    • This function initiates peer discovery for the passed in lpni and returns LNET_DC_WAIT{{
    • It is called when we want to discover a peer on first use.
    • It is called when we want to discover a gateway on first use.
  • lnet_handle_find_routed_path()
    • Call lnet_find_routed_locked() to find a gateway
    • If the gateway has not been discovered yet. Then discover it.
    • Increment the sequence number on the route only if the route is going to be used.
      • This helps in ensuring that the route sequence numbers remain sane.
  • lnet_find_route_locked()
    • returns the route to use use_route and the previous route prev_route
    • It not longer increments the sequence number of the route since finding the route doesn't equate to using the route
      • Incrementing the route sequence number is delegated to the calling function.
https://review.whamcloud.com/#/c/33183/
Code Block
LU-11298 lnet: use peer for gateway

The routing code uses peer_ni for a gateway. However with Mulit-Rail
a gateway could have multiple interfaces on several different
networks. Instead of using a single peer_ni as the gateway we should
be using the peer and let the MR selection code select the best
peer_ni to send to.

This patch moves the gateway from peer to peer_ni. Much of the
code needs to be rewritten in the following patches to account
for that change. This patch disables the routing features by
disabling the code to add/delete routes.

Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: Ia7dab552268c4a7fbd7b88122b9a95363d155fd7

The routing code will change quiet a bit so this patch removes most of the current routing code and then reintroduces it later.

This patch concentrates on switching the gateway from using lnet_peer_ni to using lnet_peer.

The design decision here is that a gateway is a node where LNet is started with the routing feature enabled. A gateway node can have multiple interfaces. In order to align routing with Multi-Rail, then the code should be first selecting a gateway peer, then using multi-rail to select the best peer_ni on that gateway to use.

The following functions are removed in this patch and will be introduced in later patches

Code Block
lnet_is_route_alive()
lnet_rtr_addref_locked()
lnet_rtr_decref_locked()
lnet_shuffle_seed()
lnet_add_route_to_rnet()
lnet_add_route() # the bulk of the code is removed
lnet_check_routes() # the bulk of the code is removed
lnet_del_route() # the bulk of the code is removed
lnet_parse_rc_info() # the bulk of the code is removed
lnet_destroy_rc_data()
lnet_update_rc_data_locked()
lnet_router_check_interval()
lnet_ping_router_locked()
lnet_prune_rc_data()
lnet_compare_peers()

Key fields are moved from lnet_peer_ni to lnet_peer or deleted including:

Code Block
lpni_rtrq # moved
lpni_rtr_list # moved
lpni_ping_notsent # deleted
lpni_ping_timestamp # deleted
lpni_ping_deadline # deleted
lpni_rtr_refcount # moved
lpni_healthy # this is a remnant code which is cleaned up
lpni_routes # moved

The lnet_route structure is changed in the following way:

Code Block
struct lnet_peer *lr_gateway # this is now lnet_peer instead of lnet_peer_ni
__u32 lr_lnet

It is no longer possible to determine the local network of the route by simply looking at the gateway peer, since the peer can have multiple interfaces on different networks. Therefore the route now must define the local network and remote network. This way we are able to select and compare routes properly.

The rest of the changes concentrate on removing the use of lnet_peer_ni as the gateway and replacing it with lnet_peer

In lib-move.c there are changes in both lnet_post_routed_recv_locked() and lnet_return_rx_credits_locked()

lnet_find_route_locked() is marked as "to be implemented". As a result lnet_handle_find_routed_path() which calls lnet_find_route_locked() is also incomplete due to removal of routing functionality. There are changes there, but the changes are mainly to avoid compilation problem. It will be re-implemented in a later patch.

Routing is disabled with this patch.

https://review.whamcloud.com/#/c/33184/
Code Block
LU-11299 lnet: lnet_add/del_route()

Reimplemented lnet_add_route() and lnet_del_route() to use
the peer instead of the peer_ni.

Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: I3734098a81ab18d1d74220c691d96a9b9817e6da

This patch re-implements the following functions, which now use lnet_peer instead of lnet_peer_ni for the gateway

Code Block
lnet_rtr_addref_locked()
lnet_rtr_decref_locked()
lnet_shuffle_seed()
lnet_add_route_to_rnet()
lnet_add_route()
lnet_del_route_from_rnet()
lnet_del_route()
https://review.whamcloud.com/#/c/33448/
Code Block
LU-11551 lnet: Do not allow deleting of router nis
    
Check the peer before deleting a peer_ni. If it's a router then do
not allow deletion of the peer-ni.
    
Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: I372052b4e9b5af3a8f18a49676fc60b4c8077cbd
Add a check before deleting the peer_ni in lnet_del_peer_ni()
https://review.whamcloud.com/#/c/33449/
Code Block
LU-11300 lnet: router sensitivity
    
Introduce the lnet_router_sensitivity module parameter to control
the sensitivity of routers to failures. It defaults to 100% which
means a router interface needs to be fully healthy in order to be
used.
    
Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: I3e9333033f049918c1cdca58a72604c71884acbe
This patch introduces the router_sensitivity_percentage module parameter
https://review.whamcloud.com/#/c/33455/
Code Block
LU-11300 lnet: configure lnet_router_sensitivity
    
Allow the configuration of lnet_router_sensitivity from the user space utility lnetctl
    
Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: I715059580d3d2d443432a8b550d4cdafc9f9f632
This patch allows setting the router sensitivity from lnetctl
https://review.whamcloud.com/#/c/33450/
Code Block
LU-11300 lnet: cache ni status

When processing the data in the PUSH or the REPLY make sure to cache
the ns_status. This is the status of the peer_ni as reported by the
peer itself.
    
Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: I14de2460f578fb7f47d329a97b8833f49c569b74

This patch caches the ns_status reported in the PING reply for a GET sent as part of discovery

  • lnet_peer_merge_data()
    • When processing the data in the PUSH or the REPLY make sure to cache the ns_status. This is the status of the peer_ni as reported by the peer itself.
https://review.whamcloud.com/#/c/33451/
Code Block
LU-11300 lnet: Cache the routing feature
    
When processing a REPLY or a PUSH for a discovery cache the
whether the routing feature is enabled or disabled as
reported by the peer.
    
Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: I69bd41fade196773af0e1004c2e7fff2fb91392d

This patch caches the routing feature status (enabled or disabled) in the REPLY or the PUSH processed by lnet_peer_merge_data().

This is important for later patches which check if the peer is a router or not.

https://review.whamcloud.com/#/c/33186/
Code Block
LU-11300 lnet: peer aliveness

Peer NI aliveness is now solely dependent on the health
infrastructure. With the addition of router_sensitivity_percentage,
peer NI is considered dead if its health drops below the percentage
specified of the total health. Setting the percentage to 100% means
that a peer_ni is considered dead if it's interface is less than
fully healthy.

Removed obsolete code that queries the peer NI every second since
the health infrastructure introduces the recovery mechanism which
is designed to recover the health of peer NIs.

Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: I506060fbb66c74295808891b689d7d634dc69284

  • lnet_is_peer_ni_alive()
    • Determines if the peer is alive if it's health value >= MAX * router sensitvity percentage
  • and
    • the status reported in the ping is UP
  • lnet_peer_alive_locked() - changes
  • Set the lpni_ns_status = LNET_NI_STATUS_UP when we receove a message from that peer.

Code which calls this function is updated.

router_proc.c cleanup

https://review.whamcloud.com/#/c/33185/
Code Block
LU-11300 lnet: router aliveness

A route is considered alive if the gateway is able to route
messages from the local to the remote net. That means that
at least one of the network interfaces on the remote net of
the gateway is viable.

Introduced the concept of sensitivity percentage. This defaults
to 100%. It holds a dual meaning:
1. A route is considered alive if at least one of the its interfaces'
health is >= LNET_MAX_HEALTH_VALUE * router_sensitivity_percentage
100 means at least one interface has to be 100% healthy
2. On a router consider a peer_ni dead if its health is not at least
LNET_MAX_HEALTH_VALUE * router_sensitivity_percentage.
100% means the interface has to be 100% healthy.

Re-implemented lnet_notify() to decrement the health of the
peer interface if the LND reports a failure on that peer.

Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: Ie97561fb70bf6a558bc90fa9266a6ba38fa3d293

NOTE: break up this patch into a patch which introduces the router_sensitvity_percentage and a patch which uses the value. The changes which are in lib-msg.c do not belong to this patch. Need a separate patch for it.

NOTE: This should be titled "route aliveness"

This patch introduces a new concept on how to determine that a route is alive. A route is alive if the following two conditions are met:

  1. At least one local interface to the route is healthy
  2. At least one remote interface of the route is up

The health value of a router remote interface will always be set to MAX because we do not send to it directly, therefore we never decrement its health value. The way we know if it's up or down is when we discover it, the router response with the status of the interface which we cache and use to determine the status of the remote interface.

  • lnet_is_gateway_net_alive()
    • A net of the gateway is alive if it has at least one alive ni
https://review.whamcloud.com/#/c/33187/
Code Block
LU-11299 lnet: Cleanup rcd
    
Cleanup all code pertaining to rcd, as routing code will use
discovery going forward and there will be no need to keep its own
pinging code.
    
Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: If31caa3b5703df40b6ae0f758f2fe764991aa4f3
Cleans up the legacy code which handled router pinging
https://review.whamcloud.com/#/c/33453/
Code Block
LU-11299 lnet: modify lnd notification mechanism
    
LND notifies when a peer is up or down. If a it notifies
LNet that the peer is up and sets the "reset" flag to true
then this indicates to LNet that the LND knows about the health
of the peer and is telling LNet that the peer is fully healthy.
LNet will set the health value of the peer to maximum, otherwise
it will increment the health by one.
    
If the LND notifies the LNet that the peer is down, LNet will
decrement the health of the peer by sensitivity value configured.
    
LNet then turns around and rechecks the peer aliveness and if its
dead it'll notify the LND. This code is only used by the socklnd
because it needs to teardown connections. This is in keeping with
the original funcionality.
    
Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: Ifa614405fb0c2cd4f6bcb1a2a97e856320eb6cbe

Updates

lnet_notify()

lnet_set_healthv()

lnet_notify_peer_down()

gni changes

o2iblnd changes

socklnd changes

https://review.whamcloud.com/#/c/33454/
Code Block
LU-11299 lnet: use discovery for routing

Instead re-inventing the wheel, routing now uses discovery.
Everyone router interval the router is discovered. This will
update the router information locally and will serve to let the 
router know that the peer is alive.

Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: I211bf15af0b0a5d50f9e2a69a385419a1dd5096

lnet_consolidate_routes_locked()

lnet_peer_get_ni_locked()

lnet_check_routers()

lnet_router_discovery_complete()

  • Called from lnet_peer_discovery_complete() for peers which are routers
  • It checks the discovery completion code and if it's an error sets all the lpni→lpni_ns_status = LNET_NI_STATUS_DOWN
https://review.whamcloud.com/#/c/33188/
Code Block
LU-11378 lnet: MR aware gateway selection

When selecting a route use the Multi-Rail Selection algorithm to
select the best available peer_ni of the best route. The selected
peer_ni can then be used to send the message or to discover it
if the gateway peer needs discovering.

Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: I376af57611591eed2eb1edb80a1b3a68b5aefd19
This patch modifies lib-move.c to properly select the gateway and then the gateway peer_ni to send to.
https://review.whamcloud.com/#/c/33298/
Code Block
LU-11300 lnet: consider router_check_interval

Consider router_check_interval when waking up the monitor thread,
to make sure you wakeup the monitor thread at the earliest possible
time.

Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: Ibc4b53886b59a9bc174a29d0da711ac77db3a62c

The monitor thread wakes up the minimum of

  • lnet_recovery_interval
  • lnet_transaction_timeout /2
  • router_check_interval

This patch introduces the router_check_interval for consideration in the monitor thread wake up algorithm

https://review.whamcloud.com/#/c/33300/
Code Block
LU-11475 lnet: allow deleting router primary_nid

Discovery doesn't allow deleting a primary_nid of a peer. This
is necessary because upper layers only know to reach the peer by
using the primary_nid. For routers this is not the case. So
if a router changes its interfaces and comes back up again, the
peer_ni should be adjusted.

Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: I9da056172f35a5f15eed5ba0e02fcb37ac414c54
  • Add a new bit to the peer state: LNET_PEER_RTR_NI_FORCE_DEL
  • add a force parameter to lnet-peer_ni_del_locked() this will be set to true in lnet_peer_merge_data()

    • for routers it's okay to delete the primary_nid because the upper layers don't really rely on it. So if we're being told that the router changed its primary_nid then it's okay to delete it.

https://review.whamcloud.com/#/c/33301/
Code Block
LU-11477 lnet: handle health for incoming messages

In case of routers (as well as for the general case) it's important to
update the health of the ni/lpni for incoming messages. For an lpni
specifically when we receive a message is when we know that the lpni
is up.

A percentage router health is required in order to send a message to a
gateway. That defaults to 100, meaning that a router interface has to
be absolutely healthy in order to send to it. This matches the current
behavior. So if a router interface goes down an its health goes down
significantly, but then it comes back up again; either we receive a
message from it or we discover it and get a reply, then in order to
start using that router interface again we have to boost its health
all the way up to maximum.

This behavior is special cased for routers.

Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: Ida6c23f95dbef56c2e6ed7b6d03743939d8b30a0

Most of the modifications are in lnet_health_check()

  • handle incoming messages
  • Only increment the health of a peer_ni if we receive a message from it. Sending a message successfully doesn't mean the peer_ni is actually alive.
  • For router case boost the health of the peer_ni of the router to max if we receive a message from it.
    • We need to do that because if we increment it by 1 only and the percentage sensitivity is set to 100 we will not use that router interface for routing. This is to keep current routing behavior.
https://review.whamcloud.com/#/c/33304/
Code Block
LU-11478 lnet: misleading discovery seqno.

There is a sequence number used when sending discovery messages. This
sequence number is intended to detect stale messages. However it
could be misleading if the peer reboots. In this case the peer's
sequence number will reset. The node will think that all information
being sent to it is stale, while in reality the peer might've
changed configuration.

There is no reliable why to know whether a peer rebooted, so we'll
always assume that the messages we're receiving are valid. So we'll
operate on first come first serve basis.

Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: I421a00e47bc93ee60fa37c648d6d9a726d9def9c

Need to pass this by Olaf Weber

https://review.whamcloud.com/#/c/33305/
Code Block
LU-11470 lnet: drop all rule

Add a rule to drop all messages arriving on a specific interface.
This is useful for simulating failures on a specific router interface.

Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: Ic69f683fb2caf7a69a1d85428878c89b7b1ee3ad

For testing routers we want to be able to add a rule on the router to drop all messages arriving on that router interface from anywhere. This way we can simulate a router interface down scenario. Problem is the source and destination in the router case are not the router NID. So the rule specifies the local NID of the router.

If the local  nid is not specific then it default to LNET_NID_ANY. Unlike source and destination it is mandatory. specifying NID any allows the drop rule to match messages in the absence of a specified local_nid

drop all field is added which can be set from command line.



https://review.whamcloud.com/#/c/33620/3
Code Block
LU-11641 lnet: handle discovery off

When discovery is turned off locally or when the peer either has
discovery off or doesn't support MR at all then degrade discovery
behavior to a standard ping. This will allow routers to continue
using discovery mechanism even if it's turned off.

Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: I7f0829d37cbff2bf9e41de251efa715fc4c97e5d

The original discovery behavior was that if you turn off discovery then it doesn't send a PING or a PUSH. However, this causes a problem for routers. One example is Cray need to turn off discovery because it interferes with their routing setup. To handle this we changed the behavior for discovery off case. Discovery will always PING when requested. When the PING REPLY comes it'll update the existing peer_nis but not add or move peers_nis among peers. This is needed to update the router peer_ni statuses on the nodes.

Most of the changes are spread through peer.c to change what we do when discovery on the peer is off or when it is turned off locally.

https://review.whamcloud.com/#/c/33634/2
Code Block
LU-11297 lnet: handle router health off

Routing infrastructure depends on health infrastructure to manage
route status. However, health can be turned off. Therefore, we need
to enable health for gateways in order to monitor them properly.
Each peer now has its own health sensitivity. When adding a route
the gateway's health sensitivity can be explicitly set from lnetctl
or if not specified then it'll default to 1, thereby turning health
on for that gateway, allowing peer NI recovery if there is a failure.

Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: Ibae33d595e97d0eec432ae8f5d51898ce0776f01
The way router health is decided is via the health mechanism. This presents a problem since health is turned off by default. This patch enables it for all routers.
https://review.whamcloud.com/#/c/33635/2
Code Block
LU-11297 lnet: set gw sensitivity from lnetctl

Allow an optional parameter from the:
lnetctl route add
command to set the health sensitivity of the gateway
lnetctl route add --net <net> --gateway <gw> --sensitivity <value>

Test-Parameters: forbuildonly
Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: Iee120c78a41b79da6ab6bdf1560f558df89233e2
Allow configuring router sensitivity value when adding a route. It's an optional parameter and defaults to 1 if not specified.
https://review.whamcloud.com/#/c/33651/
Code Block
LU-11664 lnet: push router interface updates

A router can bring up/down its interfaces if it hasn't received any
messages on that interface for a configurable period
(alive_router_ping_timeout). When this even occures the router can now
push its status change to the peers it's talking to in order to inform
them of the change in its status. This will allow the router users to
handle asym router failures quicker.

Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: I9530ed7d9bc0a86edc43e3f610cc943f1732dcfd
To allow for faster recovery if the router interfaces go down because it doesn't receive a ping, a push with the updated information is pushed to all the peers. This allows peers to stop using these routers if asym router failure is enabled (it is by default).