Re: [RFC PATCH net-next] qfq: handle the case that front slot is empty

Il giorno 28/ott/2012, alle ore 13:45, Cong Wang ha scritto:

On Fri, 2012-10-26 at 18:51 +0200, Paolo Valente wrote:

quoted

Here is a possible patch. Could you please give me a feedback?

If this patch actually works, there are some issues related to it that I would
like to point out after your (and/or anyone else's) tests.

With this patch applied, I can't reproduce the crash any more.

Thank you for the test.

And the following messages appear in the log:

[  430.956400] qfq: increasing class max_pkt_size from 1514 to 26130
[  430.981887] qfq: increasing class max_pkt_size from 26130 to 31922
[  431.011844] qfq: increasing class max_pkt_size from 1514 to 2962
[  431.060129] qfq: increasing class max_pkt_size from 31922 to 46402
[  431.096241] qfq: increasing class max_pkt_size from 2962 to 4410
[  431.123689] qfq: increasing class max_pkt_size from 4410 to 18890
[  431.134624] qfq: increasing class max_pkt_size from 18890 to 20338
[  431.149573] qfq: increasing class max_pkt_size from 20338 to 39162
[  431.328157] qfq: increasing class max_pkt_size from 46402 to 56538
[  431.339731] qfq: increasing class max_pkt_size from 56538 to 65226
[  431.363699] qfq: increasing class max_pkt_size from 39162 to 44954
[  431.376303] qfq: increasing class max_pkt_size from 44954 to 46402
[  431.482082] qfq: increasing class max_pkt_size from 46402 to 62330
[  431.520339] qfq: increasing class max_pkt_size from 62330 to 65226

WIth each of these messages, qfq now informs the user that he/she has chosen a too low max_pkt_size with respect to the actual size of the last-arrived packet, and that the max_pkt_size has been increased automatically to handle this larger packet correctly. In the log that you report, the large packets that cause qfq to increase the max_pkt_size are due to TSO/GSO.

Any feedback about these notifications is welcome. Meanwhile I will prepare a description of the patch and of the limitations it imposes.