- Background: ixgbe doesn't use common MCLGET() interface and use the
driver specific cluster allocation mechanism (jcl). The cluster is
pre-allocated with a fixed number and the current number per queue
is num_rx_desc * 2 (2048*2=4096). It's too small. It also has a problem
that the max length of the pcq which is used in the TX path is big
(4096). Example:
100M <----- [ixg0 ixg1] <----- 1G
2048 TX descs <--- 4096 pcqs <---- 2048 RX descs
If a machine forwards a traffic from 1G interface to 100M interface,
It would require 2048+4096+2048=8192 descriptors, but the current number
is 2048*2=4096. It's too small. Even if the both interface's link speed
is the same and only small number of packet is queued in the pcq, 4096
jcl is small because 2048(RX)+TX(2048)=4096. If jcl is exhausted, not only
forwarding from ixg1 to ixg0 is dropped, but also another forwarding path
from ixg1 to another interface(e.g. wm0) is also dropped. Sockets also
queue packets, so if a lot of sockets are used and/or a socket buffer
size is changed to bigger one, it'll also become a problem. If the jcl
is exhausted, evcnt(9) counter "ixgX qY Rx no jumbo mbuf" is incremented.
Example:
vmstat -ev | grep ixg1 | grep "no jumbo"
ixg1 q0 Rx no jumbo mbuf 0 0 misc
ixg1 q1 Rx no jumbo mbuf 0 0 misc
ixg1 q2 Rx no jumbo mbuf 141326 0 misc
ixg1 q3 Rx no jumbo mbuf 0 0 misc
- To solve this problem:
- Add new config parameter IXGBE_JCLNUM_MULTI and set the default to 3
(2048 * 3). The minimum number is 2. The total number of jcl per queue
is available with hw.ixgN.num_jcl_per_queue sysctl.
- Reduce the max length of the pcq() which is used in the TX path from
4096 to 2048.
- Reviewed by knakahara@ and ozaki-r@.
- TODO: Use MCLGET().
It may have been a clever trick to use the same memory layout for struct
templ and a string pointer, but it's not worth the extra comment and
difficulty in understanding the code.
No functional change.
When indent runs in filter mode, it may output messages to stderr.
Allow tests with non-empty expected stderr.
In the ATF output, the filename 'output_file.parsed' was not helpful for
casual readers of diff output since they expect the filenames to be
meaningful. Embed the name of the test case in that filename.
Fix quoting of the shell variables.
Remove the repetition of the regular expression to clean up the test
files.
ryo