K 10
svn:author
V 5
markj
K 8
svn:date
V 27
2018-04-11T14:38:56.141158Z
K 7
svn:log
V 1051
Let the driver specify the required number of netdump clusters.

In most cases, NETDUMP_MAX_IN_FLIGHT (64) is sufficient. However, the
iflib and cxgb rx buffer allocation schemes have an edge case which
make this insufficient. Both drivers allocate mbufs and clusters
separately and keep them in separate arrays. If a received packet can
fit in the mbuf, the cluster at the corresponding index is not consumed
and will simply be left alone until the consumer wraps around. netdump
ACKs fit in an mbuf's inline buffer, so the clusters are generally not
used at all. However, if the drivers fail to allocate a cluster for a
given producer index, they will not replenish the mbuf either, so unless
there are enough pre-allocated clusters for the whole ring (which almost
certainly has more than 64 entries), we quickly stop being able to
receive packets.

For these drivers, be sure to pre-allocate enough clusters to avoid
the problem. This is quite wasteful but is the simplest path forward
for now. Other drivers continue to pre-allocate 64 clusters.

END