Add a nexthop hashing api for only hashing on word-sized
attributes. Calling the jhash/jhash2 function is quite slow
in scaled envrionments but sometimes you do need a more granular hash.
The tradeoff here is that hashtable buckets using this hash
might be more full.
Signed-off-by: Stephen Worley <sworley@cumulusnetworks.com>
In the nexthop hashing function, lets reduce the hash calls as
much as possible. So, reduce the onlink and infindex to one
call to jhash_2words().
Signed-off-by: Stephen Worley <sworley@cumulusnetworks.com>
Optimize the fib and notified nexthop group comparison algorithm
to assume ordering. There were some pretty serious performance hits with
this on high ecmp routes.
Signed-off-by: Stephen Worley <sworley@cumulusnetworks.com>
Separate nexthop_group_equal() into two versions. One
that compares verses recurisvely resolved nexthops and
one that doesn't.
Signed-off-by: Stephen Worley <sworley@cumulusnetworks.com>
We were waiting until install time to mark nexthops as duplicate.
Since they are immutable now and re-used, move this marking into
when they are actually created to save a bunch of cycles.
Signed-off-by: Stephen Worley <sworley@cumulusnetworks.com>
When hashing a nexthop, shove all the nexthop g_addr data together
and pass it as one call to jhash2() to optimize a bit better.
Signed-off-by: Stephen Worley <sworley@cumulusnetworks.com>
Speed up nexthop_group_equal() by making it assume the
groups it has been passed are ordered. This should always
be the case.
Signed-off-by: Stephen Worley <sworley@cumulusnetworks.com>
We should hash nexthops on the onlink flag since that is a
descriptor of the nexthop and not a status of it.
Signed-off-by: Stephen Worley <sworley@cumulusnetworks.com>
Include resolved nexthops when hashing a nexthop
group but provide an API that allows you to non-recursively
hash as well.
Signed-off-by: Stephen Worley <sworley@cumulusnetworks.com>
Add the ability to recursively resolve nexthop group hash entries
and resolve them when sending to the kernel.
When copying over nexthops into an NHE, copy resolved info as well.
Signed-off-by: Stephen Worley <sworley@cumulusnetworks.com>
We will use a nhe context for dataplane interaction with
nextho group hash entries.
New nhe's from the kernel will be put into a group array
if they are a group and queued on the rib metaq to be processed
later.
New nhe's sent to the kernel will be set on the dataplane context
with approprate ID's in the group array if needed.
Signed-off-by: Stephen Worley <sworley@cumulusnetworks.com>
Add a function to check whether nexthop groups
are equivalent. It does not care about ordering.
Also, set any functions that it relies on to take
const vars as well.
Signed-off-by: Stephen Worley <sworley@cumulusnetworks.com>
Guard the libyang debug messages under this command so that only
people interested on those messages will see them.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
getrusage, in a heavily stressed system, can account for
signficant running time due to process switching to the kernel.
Allow the end-operator to specify `--disable-cpu-time` to
avoid this call. Additionally we cause `show thread cpu` to
not show up if this is selected.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
Add a new function getsockopt_so_recvbuf which tells you the
operating systems receive buffer size.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
Currently libyang logs errors only (LY_LLERR by default), independent of
FRR's log level. This commit lets libyang log everything including all
sorts of debug logs (when libyang is built in 'Debug' mode). FRR's
logging infrastructure filters logs out according to the configured log
level.
There is a very small performance overhead involved, even when libyang
is build in 'Release' mode. This overhead is mainly affecting config
processing and barely measurable being around 0-3% of the processing
time without this change.
Signed-off-by: Sascha Kattelmann <sascha@netdef.org>
With commit: a9ff90c41b
the vrf_id_t was changed from a uint16_t to a uint32_t
Zebra tracked the last command sent to it's peer via peeking
into the data it was sending to each client ( since we had
lost the idea of what the command was when it was time to track
the data ).
Add a define to track this and add a bit of verbiage
to the code to allow us to notice when we screw with
the header again so that this is just fixed correctly
when it happens again.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
When POLLNVAL is received for a FD then that FD is removed from the
pfd array and also array is rearranged using memmove. When memmove
is used then unused index are not cleanedup. When a new FD takes
up that index then it ends up using stale events without any handler
set for the same.
Signed-off-by: Santosh P K <sapk@vmware.com>
The dnode member of the nb_config structure can be null on
daemons that don't implement any YANG module. As such, update
the nb_cli_show_config_prepare() function to always check if the
libyang data node that is going to be displayed is null or not
before operating on it.
This fixes the following warning (introduced by commit 5e6a9350c1):
libyang: Invalid arguments (lyd_schema_sort())
Reported-by: Donald Sharp <sharpd@cumulusnetworks.com>
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
This change fixes a static analyzer warning and should also make us
safer when using this function. At the moment the code that triggered
the warning is the only one that uses this function.
Passing anything other than `struct prefix` to `str2prefix` function is
dangerous, because the structure might be smaller than expected and we
might have an buffer overflow.
Signed-off-by: Rafael Zalamena <rzalamena@opensourcerouting.org>
Remove the xpath field from the nb_config_cb structure in order
to reduce its size. This allows the northbound to spend less time
allocating memory during the processing of large configuration
transactions.
To make this work, use yang_dnode_get_path() to obtain the xpath
from the dnode field whenever necessary.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
Commit 6b5d6e2dbc changed how we order configuration callbacks
and introduced a regression in the processing of the 'apply_finish'
callbacks. Fix this.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
Load the startup configuration directly into the CLI shared candidate
configuration instead of loading it into a private candidate
configuration. This way we don't need to initialize the shared
candidate separately later as a copy of the running configuration,
which is a potentially expensive operation.
Also, make the northbound process SIGHUP correctly even when --tcli
is not used.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
nb_candidate_edit() was calling both the lyd_schema_sort() and
lyd_validate() functions whenever a new node was added to the
candidate configuration. This was done to ensure the candidate
is always ready to be displayed correctly (libyang only creates
default child nodes during the validation process, and data nodes
aren't guaranteed to be ordered by default).
The problem is that the two aforementioned functions are too
expensive to be called in the northbound hot path. Instead, it makes
more sense to call them only before displaying the configuration
(in which case a recursive sort needs to be done). Introduce the
nb_cli_show_config_prepare() to achieve that purpose.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
The nb_cli_apply_changes() function was creating a full copy of
the candidate configuration before editing it. This excerpt from
the northbond documentation explains why this was being done:
"NOTE: the nb_cli_cfg_change() function clones the candidate
configuration before actually editing it. This way, if any error
happens during the editing, the original candidate is restored to
avoid inconsistencies. Either all changes from the configuration
command are performed successfully or none are. It's like a
mini-transaction but happening on the candidate configuration
(thus the northbound callbacks are not involved)".
The problem is that this kind of error handling is just too
expensive. A command should never fail to edit the candidate
configuration unless there's a bug in the code (e.g. when the
CLI wrapper command passes an integer value that YANG rejects due
to a "range" statement). In such cases, a command might fail to
be applied or applied only partially if it edits multiple YANG
nodes. When that happens, just log an error to make the operator
aware of the problem, but otherwise ignore it instead of rejecting
the command and restoring the candidate to its previous state. We
shouldn't add an extreme overhead to the northbound CLI client only
to handle errors that should never happen in practice.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
There's no need to check for removed configuration objects in the
following two cases:
* A private candidate configuration is being edited;
* The startup configuration is being read.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
