On Thu, Aug 11, 2016 at 9:20 AM, Jamal Hadi Salim <j...@mojatatu.com> wrote: > On 16-08-10 04:06 PM, Cong Wang wrote: >> >> On Wed, Aug 10, 2016 at 7:34 AM, Jamal Hadi Salim <j...@mojatatu.com> >> wrote: >>> >>> On 16-08-08 04:46 PM, Cong Wang wrote: > > >>> tcf_exts_exec() is the culprit - and conversion to from flexarray >>> to linked list in the fast problem to be specific. >> >> >> Ah, this reminds me that I don't have to use flex_array, initially >> I thought the tcf_exts could hold as many actions as it wants, >> but actually there is a upper bound, TCA_ACT_MAX_PRIO. >> IOW, a regular dynamic array is just enough here. >> > > Yes, a regular array would be enough. > > >> I just replaced the flex_array with a regular one, it works fine >> for me too, at least no crash with all of my test cases. >> > > Ok, I did a quick look at your patch - I still see you converting > from array to list. I mean get tcf_exts_exec() to take an array > and walk it. That will restore the perf numbers to the same level.
Makes sense! I just did this change as you suggest. > >> Please try v2, since you have more test cases that I do. >> Or it would be great if you can share your test cases with >> me or us. >> >> Be patient, every big change could have regression. :) >> > > No problem Cong - except we have a kernel that crashes right now. > BTW: I just thought of another test which uses a different code > path > # add a policer rule > sudo $TC actions add action police rate 1kbit burst 90k drop > #dump rules.. > sudo $TC -s actions ls action police > Passed: [root@localhost ~]# tc actions add action police rate 1kbit burst 90k drop [root@localhost ~]# tc actions ls action police action order 0: police 0x1 rate 1000bit burst 23440b mtu 2Kb action drop overhead 0b ref 1 bind 0 [root@localhost ~]# tc -s actions ls action police action order 0: police 0x1 rate 1000bit burst 23440b mtu 2Kb action drop overhead 0b ref 1 bind 0 Action statistics: Sent 0 bytes 0 pkt (dropped 0, overlimits 0 requeues 0) backlog 0b 0p requeues 0 Thanks.
