The cma_acquire_dev function was changed by commit 3c86aa70bf
("RDMA/cm: Add RDMA CM support for IBoE devices") to use find_gid_port()
because multiport devices might have either IB or IBoE formatted gids.
The old function assumed that all ports on the same device used the
same GID format.
However, when it was changed to use find_gid_port(), we inadvertently
lost usage of the GID cache. This turned out to be a very costly
change. In our testing, each iteration through each index of the GID
table takes roughly 35us. When you have multiple devices in a system,
and the GID you are looking for is on one of the later devices, the
code loops through all of the GID indexes on all of the early devices
before it finally succeeds on the target device. This pathological
search behavior combined with 35us per GID table index retrieval
results in results such as the following from the cmtime application
that's part of the latest librdmacm git repo:
ib1:
step total ms max ms min us us / conn
create id : 29.42 0.04 1.00 2.94
bind addr : 186705.66 19.00 18556.00 18670.57
resolve addr : 41.93 9.68 619.00 4.19
resolve route: 486.93 0.48 101.00 48.69
create qp : 4021.95 6.18 330.00 402.20
connect : 68350.39 68588.17 24632.00 6835.04
disconnect : 1460.43 252.65-1862269.00 146.04
destroy : 41.16 0.04 2.00 4.12
ib0:
step total ms max ms min us us / conn
create id : 28.61 0.68 1.00 2.86
bind addr : 2178.86 2.95 201.00 217.89
resolve addr : 51.26 16.85 845.00 5.13
resolve route: 620.08 0.43 92.00 62.01
create qp : 3344.40 6.36 273.00 334.44
connect : 6435.99 6368.53 7844.00 643.60
disconnect : 5095.38 321.90 757.00 509.54
destroy : 37.13 0.02 2.00 3.71
Clearly, both the bind address and connect operations suffer
a huge penalty for being anything other than the default
GID on the first port in the system.
After applying this patch, the numbers now look like this:
ib1:
step total ms max ms min us us / conn
create id : 30.15 0.03 1.00 3.01
bind addr : 80.27 0.04 7.00 8.03
resolve addr : 43.02 13.53 589.00 4.30
resolve route: 482.90 0.45 100.00 48.29
create qp : 3986.55 5.80 330.00 398.66
connect : 7141.53 7051.29 5005.00 714.15
disconnect : 5038.85 193.63 918.00 503.88
destroy : 37.02 0.04 2.00 3.70
ib0:
step total ms max ms min us us / conn
create id : 34.27 0.05 1.00 3.43
bind addr : 26.45 0.04 1.00 2.64
resolve addr : 38.25 10.54 760.00 3.82
resolve route: 604.79 0.43 97.00 60.48
create qp : 3314.95 6.34 273.00 331.49
connect : 12399.26 12351.10 8609.00 1239.93
disconnect : 5096.76 270.72 1015.00 509.68
destroy : 37.10 0.03 2.00 3.71
It's worth noting that we still suffer a bit of a penalty on
connect to the wrong device, but the penalty is much less than
it used to be. Follow on patches deal with this penalty.
Many thanks to Neil Horman for helping to track the source of
slow function that allowed us to track down the fact that
the original patch I mentioned above backed out cache usage
and identify just how much that impacted the system.
Signed-off-by: Doug Ledford <dledford@redhat.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
29f27e8477