The current design of the binding table has an unnecessary memory consuming and complex data structure. It aggregates the service range items into an array, which is expanded by a factor two every time it becomes too small to hold a new item. Furthermore, the arrays never shrink when the number of ranges diminishes. We now replace this array with an RB tree that is holding the range items as tree nodes, each range directly holding a list of bindings. This, along with a few name changes, improves both readability and volume of the code, as well as reducing memory consumption and hopefully improving cache hit rate. Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
When a 32-bit node address is generated from a 128-bit identifier, there is a risk of collisions which must be discovered and handled. We do this as follows: - We don't apply the generated address immediately to the node, but do instead initiate a 1 sec trial period to allow other cluster members to discover and handle such collisions. - During the trial period the node periodically sends out a new type of message, DSC_TRIAL_MSG, using broadcast or emulated broadcast, to all the other nodes in the cluster. - When a node is receiving such a message, it must check that the presented 32-bit identifier either is unused, or was used by the very same peer in a previous session. In both cases it accepts the request by not responding to it. - If it finds that the same node has been up before using a different address, it responds with a DSC_TRIAL_FAIL_MSG containing that address. - If it finds that the address has already been taken by some other node, it generates a new, unused address and returns it to the requester. - During the trial period the requesting node must always be prepared to accept a failure message, i.e., a message where a peer suggests a different (or equal) address to the one tried. In those cases it must apply the suggested value as trial address and restart the trial period. This algorithm ensures that in the vast majority of cases a node will have the same address before and after a reboot. If a legacy user configures the address explicitly, there will be no trial period and messages, so this protocol addition is completely backwards compatible. Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
We add a 128-bit node identity, as an alternative to the currently used 32-bit node address. For the sake of compatibility and to minimize message header changes we retain the existing 32-bit address field. When not set explicitly by the user, this field will be filled with a hash value generated from the much longer node identity, and be used as a shorthand value for the latter. We permit either the address or the identity to be set by configuration, but not both, so when the address value is set by a legacy user the corresponding 128-bit node identity is generated based on the that value. Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
The removal of an internal structure of the node address has an unwanted side effect. - Currently, if a user is sending an anycast message with destination domain 0, the tipc_namebl_translate() function will use the 'closest- first' algorithm to first look for a node local destination, and only when no such is found, will it resort to the cluster global 'round- robin' lookup algorithm. - Current users can get around this, and enforce unconditional use of global round-robin by indicating a destination as Z.0.0 or Z.C.0. - This option disappears when we make the node address flat, since the lookup algorithm has no way of recognizing this case. So, as long as there are node local destinations, the algorithm will always select one of those, and there is nothing the sender can do to change this. We solve this by eliminating the 'closest-first' option, which was never a good idea anyway, for non-legacy users, but only for those. To distinguish between legacy users and non-legacy users we introduce a new flag 'legacy_addr_format' in struct tipc_core, to be set when the user configures a legacy-style Z.C.N node address. Hence, when a legacy user indicates a zero lookup domain 'closest-first' is selected, and in all other cases we use 'round-robin'. Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
As a consequence of the previous commit we nan now eliminate zone scope related lists in the name table. We start with name_table::publ_list[3], which can now be replaced with two lists, one for node scope publications and one for cluster scope publications. Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
We rename struct tipc_server to struct tipc_topsrv. This reflect its now specialized role as topology server. Accoringly, we change or add function prefixes to make it clearer which functionality those belong to. There are no functional changes in this commit. Acked-by: Ying.Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
Preempt counter APIs have been split out, currently, hardirq.h just includes irq_enter/exit APIs which are not used by TIPC at all. So, remove the unused hardirq.h. Signed-off-by: Yang Shi <yang.s@alibaba-inc.com> Acked-by: Ying Xue <ying.xue@windriver.com> Tested-by: Ying Xue <ying.xue@windriver.com> Cc: Jon Maloy <jon.maloy@ericsson.com> Cc: "David S. Miller" <davem@davemloft.net> Signed-off-by: David S. Miller <davem@davemloft.net>
As preparation for introducing communication groups, we add the ability to issue topology subscriptions and receive topology events from kernel space. This will make it possible for group member sockets to keep track of other group members. Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: David S. Miller <davem@davemloft.net>
Make struct pernet_operations::id unsigned.
There are 2 reasons to do so:
1)
This field is really an index into an zero based array and
thus is unsigned entity. Using negative value is out-of-bound
access by definition.
2)
On x86_64 unsigned 32-bit data which are mixed with pointers
via array indexing or offsets added or subtracted to pointers
are preffered to signed 32-bit data.
"int" being used as an array index needs to be sign-extended
to 64-bit before being used.
void f(long *p, int i)
{
g(p[i]);
}
roughly translates to
movsx rsi, esi
mov rdi, [rsi+...]
call g
MOVSX is 3 byte instruction which isn't necessary if the variable is
unsigned because x86_64 is zero extending by default.
Now, there is net_generic() function which, you guessed it right, uses
"int" as an array index:
static inline void *net_generic(const struct net *net, int id)
{
...
ptr = ng->ptr[id - 1];
...
}
And this function is used a lot, so those sign extensions add up.
Patch snipes ~1730 bytes on allyesconfig kernel (without all junk
messing with code generation):
add/remove: 0/0 grow/shrink: 70/598 up/down: 396/-2126 (-1730)
Unfortunately some functions actually grow bigger.
This is a semmingly random artefact of code generation with register
allocator being used differently. gcc decides that some variable
needs to live in new r8+ registers and every access now requires REX
prefix. Or it is shifted into r12, so [r12+0] addressing mode has to be
used which is longer than [r8]
However, overall balance is in negative direction:
add/remove: 0/0 grow/shrink: 70/598 up/down: 396/-2126 (-1730)
function old new delta
nfsd4_lock 3886 3959 +73
tipc_link_build_proto_msg 1096 1140 +44
mac80211_hwsim_new_radio 2776 2808 +32
tipc_mon_rcv 1032 1058 +26
svcauth_gss_legacy_init 1413 1429 +16
tipc_bcbase_select_primary 379 392 +13
nfsd4_exchange_id 1247 1260 +13
nfsd4_setclientid_confirm 782 793 +11
...
put_client_renew_locked 494 480 -14
ip_set_sockfn_get 730 716 -14
geneve_sock_add 829 813 -16
nfsd4_sequence_done 721 703 -18
nlmclnt_lookup_host 708 686 -22
nfsd4_lockt 1085 1063 -22
nfs_get_client 1077 1050 -27
tcf_bpf_init 1106 1076 -30
nfsd4_encode_fattr 5997 5930 -67
Total: Before=154856051, After=154854321, chg -0.00%
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TIPC based clusters are by default set up with full-mesh link connectivity between all nodes. Those links are expected to provide a short failure detection time, by default set to 1500 ms. Because of this, the background load for neighbor monitoring in an N-node cluster increases with a factor N on each node, while the overall monitoring traffic through the network infrastructure increases at a ~(N * (N - 1)) rate. Experience has shown that such clusters don't scale well beyond ~100 nodes unless we significantly increase failure discovery tolerance. This commit introduces a framework and an algorithm that drastically reduces this background load, while basically maintaining the original failure detection times across the whole cluster. Using this algorithm, background load will now grow at a rate of ~(2 * sqrt(N)) per node, and at ~(2 * N * sqrt(N)) in traffic overhead. As an example, each node will now have to actively monitor 38 neighbors in a 400-node cluster, instead of as before 399. This "Overlapping Ring Supervision Algorithm" is completely distributed and employs no centralized or coordinated state. It goes as follows: - Each node makes up a linearly ascending, circular list of all its N known neighbors, based on their TIPC node identity. This algorithm must be the same on all nodes. - The node then selects the next M = sqrt(N) - 1 nodes downstream from itself in the list, and chooses to actively monitor those. This is called its "local monitoring domain". - It creates a domain record describing the monitoring domain, and piggy-backs this in the data area of all neighbor monitoring messages (LINK_PROTOCOL/STATE) leaving that node. This means that all nodes in the cluster eventually (default within 400 ms) will learn about its monitoring domain. - Whenever a node discovers a change in its local domain, e.g., a node has been added or has gone down, it creates and sends out a new version of its node record to inform all neighbors about the change. - A node receiving a domain record from anybody outside its local domain matches this against its own list (which may not look the same), and chooses to not actively monitor those members of the received domain record that are also present in its own list. Instead, it relies on indications from the direct monitoring nodes if an indirectly monitored node has gone up or down. If a node is indicated lost, the receiving node temporarily activates its own direct monitoring towards that node in order to confirm, or not, that it is actually gone. - Since each node is actively monitoring sqrt(N) downstream neighbors, each node is also actively monitored by the same number of upstream neighbors. This means that all non-direct monitoring nodes normally will receive sqrt(N) indications that a node is gone. - A major drawback with ring monitoring is how it handles failures that cause massive network partitionings. If both a lost node and all its direct monitoring neighbors are inside the lost partition, the nodes in the remaining partition will never receive indications about the loss. To overcome this, each node also chooses to actively monitor some nodes outside its local domain. Those nodes are called remote domain "heads", and are selected in such a way that no node in the cluster will be more than two direct monitoring hops away. Because of this, each node, apart from monitoring the member of its local domain, will also typically monitor sqrt(N) remote head nodes. - As an optimization, local list status, domain status and domain records are marked with a generation number. This saves senders from unnecessarily conveying unaltered domain records, and receivers from performing unneeded re-adaptations of their node monitoring list, such as re-assigning domain heads. - As a measure of caution we have added the possibility to disable the new algorithm through configuration. We do this by keeping a threshold value for the cluster size; a cluster that grows beyond this value will switch from full-mesh to ring monitoring, and vice versa when it shrinks below the value. This means that if the threshold is set to a value larger than any anticipated cluster size (default size is 32) the new algorithm is effectively disabled. A patch set for altering the threshold value and for listing the table contents will follow shortly. - This change is fully backwards compatible. Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>