On Mon, Jan 14, 2019 at 11:00 AM Vlad Buslov <vla...@mellanox.com> wrote: > Hi Cong, > > Any comments regarding benchmark results? Are you okay with > spinlock-based implementation? >
Hi, Sorry for somehow forgetting to response to your last email. I suggested you to stick to mutex not because it is better than spinlock, it is because its non-atomic context makes the code better to understand and easier to review. If mutex is really slower for you, you have to make a trade-off between performance and code complexity. And, you can also choose to use mutex as a starter here, then try to move to spinlock later as a second step. You don't have to break down RTNL to a small-scope lock and convert it to a spinlock in one patchset, splitting into two patchsets makes more sense to me and is much easier for me to review. And, unlike the RX/TX path, these control paths are not critical, I don't know if we care about the ~17% difference that much. Not saying your work is not important, I am saying you are the first one who wants to improve the performance of these control paths, so I guess perhaps others don't care much. BTW, as always, please provide a git branch for your patches, I definitely don't want to apply 17 patches manually one-by-one here. :) Thanks!
